Patent Grant
12081759
The field of the invention relates to computer-implemented methods and systems for image compression and decoding, to computer-implemented methods and systems for video compression and decoding, and to related computer-implemented training methods.
There is increasing demand from users of communications networks for images and video content. Demand is increasing not just for the number of images viewed, and for the playing time of video; demand is also increasing for higher resolution, lower distortion content, if it can be provided. This places increasing demand on communications networks, and increases their energy use, for example, which has adverse cost implications, and possible negative implications for the environment, through the increased energy use.
Although image and video content is usually transmitted over communications networks in compressed form, it is desirable to increase the compression, while preserving displayed image quality, or to increase the displayed image quality, while not increasing the amount of data that is actually transmitted across the communications networks. This would help to reduce the demands on communications networks, compared to the demands that otherwise would be made.
U.S. Ser. No. 10/373,300B1 discloses a system and method for lossy image and video compression and transmission that utilizes a neural network as a function to map a known noise image to a desired or target image, allowing the transfer only of hyperparameters of the function instead of a compressed version of the image itself. This allows the recreation of a high-quality approximation of the desired image by any system receiving the hyperparameters, provided that the receiving system possesses the same noise image and a similar neural network. The amount of data required to transfer an image of a given quality is dramatically reduced versus existing image compression technology. Being that video is simply a series of images, the application of this image compression system and method allows the transfer of video content at rates greater than previous technologies in relation to the same image quality.
U.S. Ser. No. 10/489,936B1 discloses a system and method for lossy image and video compression that utilizes a metanetwork to generate a set of hyperparameters necessary for an image encoding network to reconstruct the desired image from a given noise image.
According to a first aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of:
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein in step (vii) the output image is stored.
The method may be one wherein in step (iii), quantizing the latent representation using the first computer system to produce a quantized latent comprises quantizing the latent representation using the first computer system into a discrete set of symbols to produce a quantized latent.
The method may be one wherein in step (iv) a predefined probability distribution is used for the entropy encoding and wherein in step (vi) the predefined probability distribution is used for the entropy decoding.
The method may be one wherein in step (iv) parameters characterizing a probability distribution are calculated, wherein a probability distribution characterised by the parameters is used for the entropy encoding, and wherein in step (iv) the parameters characterizing the probability distribution are included in the bitstream, and wherein in step (vi) the probability distribution characterised by the parameters is used for the entropy decoding.
The method may be one wherein the probability distribution is a (e.g. factorized) probability distribution.
The method may be one wherein the (e.g. factorized) probability distribution is a (e.g. factorized) normal distribution, and wherein the obtained probability distribution parameters are a respective mean and standard deviation of each respective element of the quantized y latent.
The method may be one wherein the (e.g. factorized) probability distribution is a parametric (e.g. factorized) probability distribution.
The method may be one wherein the parametric (e.g. factorized) probability distribution is a continuous parametric (e.g. factorized) probability distribution.
The method may be one wherein the parametric (e.g. factorized) probability distribution is a discrete parametric (e.g. factorized) probability distribution.
The method may be one wherein the discrete parametric distribution is a Bernoulli distribution, a Rademacher distribution, a binomial distribution, a beta-binomial distribution, a degenerate distribution at x0, a discrete uniform distribution, a hypergeometric distribution, a Poisson binomial distribution, a Fisher's noncentral hypergeometric distribution, a Wallenius' noncentral hypergeometric distribution, a Benford's law, an ideal and robust soliton distributions, Conway-Maxwell-Poisson distribution, a Poisson distribution, a Skellam distribution, a beta negative binomial distribution, a Boltzmann distribution, a logarithmic (series) distribution, a negative binomial distribution, a Pascal distribution, a discrete compound Poisson distribution, or a parabolic fractal distribution.
The method may be one wherein parameters included in the parametric (e.g. factorized) probability distribution include shape, asymmetry, skewness and/or any higher moment parameters.
The method may be one wherein the parametric (e.g. factorized) probability distribution is a normal distribution, a Laplace distribution, a Cauchy distribution, a Logistic distribution, a Student's t distribution, a Gumbel distribution, an Asymmetric Laplace distribution, a skew normal distribution, an exponential power distribution, a Johnson's SU distribution, a generalized normal distribution, or a generalized hyperbolic distribution.
The method may be one wherein the parametric (e.g. factorized) probability distribution is a parametric multivariate distribution.
The method may be one wherein the latent space is partitioned into chunks on which intervariable correlations are ascribed; zero correlation is prescribed for variables that are far apart and have no mutual influence, wherein the number of parameters required to model the distribution is reduced, wherein the number of parameters is determined by the partition size and therefore the extent of the locality.
The method may be one wherein the chunks can be arbitrarily partitioned into different sizes, shapes and extents.
The method may be one wherein a covariance matrix is used to characterise the parametrisation of intervariable dependences.
The method may be one wherein for a continuous probability distribution with a well-defined PDF, but lacking a well-defined or tractable formulation of its CDF, numerical integration is used through Monte Carlo (MC) or Quasi-Monte Carlo (QMC) based methods, where this can refer to factorized or to non-factorisable multivariate distributions.
The method may be one wherein a copula is used as a multivariate cumulative distribution function.
The method may be one wherein to obtain a probability density function over the latent space, the corresponding characteristic function is transformed using a Fourier Transform to obtain the probability density function.
The method may be one wherein to evaluate joint probability distributions over the pixel space, an input of the latent space into the characteristic function space is transformed, and then the given/learned characteristic function is evaluated, and the output is converted back into the joint-spatial probability space.
The method may be one wherein to incorporate multimodality into entropy modelling, a mixture model is used as a prior distribution.
The method may be one wherein to incorporate multimodality into entropy modelling, a mixture model is used as a prior distribution, comprising a weighted sum of any base (parametric or non-parametric, factorized or non-factorisable multivariate) distribution as mixture components.
The method may be one wherein the (e.g. factorized) probability distribution is a non-parametric (e.g. factorized) probability distribution.
The method may be one wherein the non-parametric (e.g. factorized) probability distribution is a histogram model, or a kernel density estimation, or a learned (e.g. factorized) cumulative density function.
The method may be one wherein the probability distribution is a non-factorisable parametric multivariate distribution.
The method may be one wherein a partitioning scheme is applied on a vector quantity, such as latent vectors or other arbitrary feature vectors, for the purpose of reducing dimensionality in multivariate modelling.
The method may be one wherein parametrisation and application of consecutive Householder reflections of orthonormal basis matrices is applied.
The method may be one wherein evaluation of probability mass of multivariate normal distributions is performed by analytically computing univariate conditional parameters from the parametrisation of the multivariate distribution.
The method may be one including use of iterative solvers.
The method may be one including use of iterative solvers to speed up computation relating to probabilistic models.
The method may be one wherein the probabilistic models include autoregressive models.
The method may be one in which an autoregressive model is an Intrapredictions, Neural Intrapredictions and block-level model, or a filter-bank model, or a parameters from Neural Networks model, or a Parameters derived from side-information model, or a latent variables model, or a temporal modelling model.
The method may be one wherein the probabilistic models include non-autoregressive models.
The method may be one in which a non-autoregressive model is a conditional probabilities from an explicit joint distribution model.
The method may be one wherein the joint distribution model is a standard multivariate distribution model.
The method may be one wherein the joint distribution model is a Markov Random Field model.
The method may be one in which a non-autoregressive model is a Generic conditional probability model, or a Dependency network.
The method may be one including use of iterative solvers.
The method may be one including use of iterative solvers to speed up inference speed of neural networks.
The method may be one including use of iterative solvers for fixed point evaluations.
The method may be one wherein a (e.g. factorized) distribution, in the form of a product of conditional distributions, is used.
The method may be one wherein a system of equations with a triangular structure is solved using an iterative solver.
The method may be one including use of iterative solvers to decrease execution time of the neural networks.
The method may be one including use of context-aware quantisation techniques by including flexible parameters in the quantisation function.
The method may be one including use of dequantisation techniques for the purpose of assimilating the quantisation residuals through the usage of context modelling or other parametric learnable neural network modules.
The method may be one wherein the first trained neural network is, or includes, an invertible neural network (INN), and wherein the second trained neural network is, or includes, an inverse of the invertible neural network.
The method may be one wherein there is provided use of FlowGAN, that is an INN-based decoder, and use of a neural encoder, for image or video compression.
The method may be one wherein normalising flow layers include one or more of additive coupling layers; multiplicative coupling layers; affine coupling layers; invertible 1×1 convolution layers.
The method may be one wherein a continuous flow is used.
The method may be one wherein a discrete flow is used.
The method may be one wherein there is provided meta-compression, where the decoder weights are compressed with a normalising flow and sent along within the bitstreams.
The method may be one wherein encoding the input image using the first trained neural network includes using one or more univariate or multivariate Padé activation units.
The method may be one wherein using the second trained neural network to produce an output image from the quantized latent includes using one or more univariate or multivariate Padé activation units.
The method may be one wherein steps (ii) to (vii) are executed wholly or partially in a frequency domain.
The method may be one wherein integral transforms to and from the frequency domain are used.
The method may be one wherein the integral transforms are Fourier Transforms, or Hartley Transforms, or Wavelet Transforms, or Chirplet Transforms, or Sine and Cosine Transforms, or Mellin Transforms, or Hankel Transforms, or Laplace Transforms.
The method may be one wherein spectral convolution is used for image compression.
The method may be one wherein spectral specific activation functions are used.
The method may be one wherein for downsampling, an input is divided into several blocks that are concatenated in a separate dimension; a convolution operation with a 1×1 kernel is then applied such that the number of channels is reduced by half, and wherein the upsampling follows a reverse and mirrored methodology.
The method may be one wherein for image decomposition, stacking is performed.
The method may be one wherein for image reconstruction, stitching is performed.
The method may be one wherein a prior distribution is imposed on the latent space, which is an entropy model, which is optimized over its assigned parameter space to match its underlying distribution, which in turn lowers encoding computational operations.
The method may be one wherein the parameter space is sufficiently flexible to properly model the latent distribution.
The method may be one wherein the first computer system is a server, e.g. a dedicated server, e.g a machine in the cloud with dedicated GPUs e.g Amazon Web Services, Microsoft Azure, etc, or any other cloud computing services.
The method may be one wherein the first computer system is a user device.
The method may be one wherein the user device is a laptop computer, desktop computer, a tablet computer or a smart phone.
The method may be one wherein the first trained neural network includes a library installed on the first computer system.
The method may be one wherein the first trained neural network is parametrized by one or several convolution matrices θ, or wherein the first trained neural network is parametrized by a set of bias parameters, non-linearity parameters, convolution kernel/matrix parameters.
The method may be one wherein the second computer system is a recipient device.
The method may be one wherein the recipient device is a laptop computer, desktop computer, a tablet computer, a smart TV or a smart phone.
The method may be one wherein the second trained neural network includes a library installed on the second computer system.
The method may be one wherein the second trained neural network is parametrized by one or several convolution matrices Ω, or wherein the first trained neural network is parametrized by a set of bias parameters, non-linearity parameters, convolution kernel/matrix parameters.
An advantage of the above is that for a fixed file size (“rate”), a reduced output image distortion may be obtained. An advantage of the above is that for a fixed output image distortion, a reduced file size (“rate”) may be obtained.
According to a second aspect of the invention, there is provided a system for lossy image or video compression, transmission and decoding, the system including a first computer system, a first trained neural network, a second computer system and a second trained neural network, wherein
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The system may be one wherein the system is configured to perform a method of any aspect of the first aspect of the invention.
According to a third aspect of the invention, there is provided a first computer system of any aspect of the second aspect of the invention.
According to a fourth aspect of the invention, there is provided a second computer system of any aspect of the second aspect of the invention.
According to a fifth aspect of the invention, there is provided a computer implemented method of training a first neural network and a second neural network, the neural networks being for use in lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that, when using the trained first neural network and the trained second neural network, for a fixed file size (“rate”), a reduced output image distortion is obtained; and for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the loss function is evaluated as a weighted sum of differences between the output image and the input training image, and the estimated bits of the quantized image latents.
The method may be one wherein the steps of the method are performed by a computer system.
The method may be one wherein the loss function is a weighted sum of a rate and a distortion.
The method may be one wherein for differentiability, actual quantisation is replaced by noise quantisation.
The method may be one wherein the noise distribution is uniform, Gaussian or Laplacian distributed, or a Cauchy distribution, a Logistic distribution, a Student's t distribution, a Gumbel distribution, an Asymmetric Laplace distribution, a skew normal distribution, an exponential power distribution, a Johnson's SU distribution, a generalized normal distribution, or a generalized hyperbolic distribution, or any commonly known univariate or multivariate distribution.
The method may be one including the steps of:
The method may be one including use of an iterative solving method.
The method may be one in which the iterative solving method is used for an autoregressive model, or for a non-autoregressive model.
The method may be one wherein an automatic differentiation package is used to backpropagate loss gradients through the calculations performed by an iterative solver.
The method may be one wherein another system is solved iteratively for the gradient.
The method may be one wherein the gradient is approximated and learned using a proxy-function, such as a neural network.
The method may be one including using a quantisation proxy.
The method may be one wherein an entropy model of a distribution with an unbiased (constant) rate loss gradient is used for quantisation.
The method may be one including use of a Laplacian entropy model.
The method may be one wherein the twin tower problem is prevented or alleviated, such as by adding a penalty term for latent values accumulating at the positions where the clustering takes place.
The method may be one wherein split quantisation is used for network training, with a combination of two quantisation proxies for the rate term and the distortion term.
The method may be one wherein noise quantisation is used for rate and STE quantisation is used for distortion.
The method may be one wherein soft-split quantisation is used for network training, with a combination of two quantisation proxies for the rate term and for the distortion term.
The method may be one wherein noise quantisation is used for rate and STE quantisation is used for distortion.
The method may be one wherein either quantisation overrides the gradients of the other.
The method may be one wherein the noise quantisation proxy overrides the gradients for the STE quantisation proxy.
The method may be one wherein QuantNet modules are used, in network training for learning a differentiable mapping mimicking true quantisation.
The method may be one wherein learned gradient mappings are used, in network training for explicitly learning the backward function of a true quantisation operation.
The method may be one wherein an associated training regime is used, to achieve such a learned mapping, using for instance a simulated annealing approach or a gradient-based approach.
The method may be one wherein discrete density models are used in network training, such as by soft-discretisation of the PDF.
The method may be one wherein context-aware quantisation techniques are used.
The method may be one wherein a parametrisation scheme is used for bin width parameters.
The method may be one wherein context-aware quantisation techniques are used in a transformed latent space, using bijective mappings.
The method may be one wherein dequantisation techniques are used for the purpose of modelling continuous probability distributions, using discrete probability models.
The method may be one wherein dequantisation techniques are used for the purpose of assimilating the quantisation residuals through the usage of context modelling or other parametric learnable neural network modules.
The method may be one including modelling of second-order effects for the minimisation of quantisation errors.
The method may be one including computing the Hessian matrix of the loss function.
The method may be one including using adaptive rounding methods to solve for the quadratic unconstrained binary optimisation problem posed by minimising the quantisation errors.
The method may be one including maximising mutual information of the input and output by modelling the difference {circumflex over (x)} minus x as noise, or as a random variable.
The method may be one wherein the input x and the noise are modelled as zero-mean independent Gaussian tensors.
The method may be one wherein the parameters of the mutual information are learned by neural networks.
The method may be one wherein an aim of the training is to force the encoder-decoder compression pipeline to maximise the mutual information between x and {circumflex over (x)}. The method may be one wherein the method of training directly maximises mutual information in a one-step training process, where the x and noise are fed into respective probability networks S and N, and the mutual information over the entire pipeline is maximised jointly.
The method may be one wherein firstly, the network S and N is trained using negative log-likelihood to learn a useful representation of parameters, and secondly, estimates of the parameters are then used to estimate the mutual information and to train the compression network, however gradients only impact the components within the compression network; components are trained separately.
The method may be one including maximising mutual information of the input and output of the compression pipeline by explicitly modelling the mutual information using a structured or unstructured bound.
The method may be one wherein the bounds include Barber & Agakov, or InfoNCE, or TUBA, or Nguyen-Wainwright-Jordan (NWJ), or Jensen-Shannon (JS), or TNCE, or BA, or MBU, or Donsker-Varadhan (DV), or IWHVI, or SIVI, or IWAE.
The method may be one including a temporal extension of mutual information that conditions the mutual information of the current input based on N past inputs.
The method may be one wherein conditioning the joint and the marginals is used based on N past data points.
The method may be one wherein maximising mutual information of the latent parameter y and a particular distribution P is a method of optimising for rate in the learnt compression pipeline.
The method may be one wherein maximising mutual information of the input and output is applied to segments of images.
The method may be one wherein encoding the input image using the first neural network includes using one or more univariate or multivariate Padé activation units.
The method may be one wherein using the second neural network to produce an output image from the quantized latent includes using one or more univariate or multivariate Padé activation units.
The method may be one wherein when back-propagating the gradient of the loss function through the second neural network and through the first neural network, parameters of the one or more univariate or multivariate Padé activation units of the first neural network are updated, and parameters of the one or more univariate or multivariate Padé activation units of the second neural network are updated.
The method may be one wherein in step (ix), the parameters of the one or more univariate or multivariate Padé activation units of the first neural network are stored, and the parameters of the one or more univariate or multivariate Padé activation units of the second neural network are stored.
An advantage of the above is that, when using the trained first neural network and the trained second neural network, for a fixed file size (“rate”), a reduced output image distortion may be obtained; and for a fixed output image distortion, a reduced file size (“rate”) may be obtained.
According to a sixth aspect of the invention, there is provided a computer program product for training a first neural network and a second neural network, the neural networks being for use in lossy image or video compression, transmission and decoding, the computer program product executable on a processor to:
The computer program product may be one wherein the loss function is evaluated as a weighted sum of differences between the output image and the input training image, and the estimated bits of the quantized image latents.
The computer program product may be executable on the processor to perform a method of any aspect of the fifth aspect of the invention.
According to a seventh aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein in step (xiii) the output image is stored.
The method may be one wherein in step (iii), quantizing the y latent representation using the first computer system to produce a quantized y latent comprises quantizing the y latent representation using the first computer system into a discrete set of symbols to produce a quantized y latent.
The method may be one wherein in step (v), quantizing the z latent representation using the first computer system to produce a quantized z latent comprises quantizing the z latent representation using the first computer system into a discrete set of symbols to produce a quantized z latent.
The method may be one wherein in step (vi) a predefined probability distribution is used for the entropy encoding of the quantized z latent and wherein in step (x) the predefined probability distribution is used for the entropy decoding to produce the quantized z latent.
The method may be one wherein in step (vi) parameters characterizing a probability distribution are calculated, wherein a probability distribution characterised by the parameters is used for the entropy encoding of the quantized z latent, and wherein in step (vi) the parameters characterizing the probability distribution are included in the second bitstream, and wherein in step (x) the probability distribution characterised by the parameters is used for the entropy decoding to produce the quantized z latent.
The method may be one wherein the (e.g. factorized) probability distribution is a (e.g. factorized) normal distribution, and wherein the obtained probability distribution parameters are a respective mean and standard deviation of each respective element of the quantized y latent.
The method may be one wherein the (e.g. factorized) probability distribution is a parametric (e.g. factorized) probability distribution.
The method may be one wherein the parametric (e.g. factorized) probability distribution is a continuous parametric (e.g. factorized) probability distribution.
The method may be one wherein the parametric (e.g. factorized) probability distribution is a discrete parametric (e.g. factorized) probability distribution.
The method may be one wherein the discrete parametric distribution is a Bernoulli distribution, a Rademacher distribution, a binomial distribution, a beta-binomial distribution, a degenerate distribution at x0, a discrete uniform distribution, a hypergeometric distribution, a Poisson binomial distribution, a Fisher's noncentral hypergeometric distribution, a Wallenius' noncentral hypergeometric distribution, a Benford's law, an ideal and robust soliton distributions, Conway-Maxwell-Poisson distribution, a Poisson distribution, a Skellam distribution, a beta negative binomial distribution, a Boltzmann distribution, a logarithmic (series) distribution, a negative binomial distribution, a Pascal distribution, a discrete compound Poisson distribution, or a parabolic fractal distribution.
The method may be one wherein parameters included in the parametric (e.g. factorized) probability distribution include shape, asymmetry and/or skewness parameters.
The method may be one wherein the parametric (e.g. factorized) probability distribution is a normal distribution, a Laplace distribution, a Cauchy distribution, a Logistic distribution, a Student's t distribution, a Gumbel distribution, an Asymmetric Laplace distribution, a skew normal distribution, an exponential power distribution, a Johnson's SU distribution, a generalized normal distribution, or a generalized hyperbolic distribution.
The method may be one wherein the parametric (e.g. factorized) probability distribution is a parametric multivariate distribution.
The method may be one wherein the latent space is partitioned into chunks on which intervariable correlations are ascribed; zero correlation is prescribed for variables that are far apart and have no mutual influence, wherein the number of parameters required to model the distribution is reduced, wherein the number of parameters is determined by the partition size and therefore the extent of the locality.
The method may be one wherein the chunks can be arbitrarily partitioned into different sizes, shapes and extents.
The method may be one wherein a covariance matrix is used to characterise the parametrisation of intervariable dependences.
The method may be one wherein for a continuous probability distribution with a well-defined PDF, but lacking a well-defined or tractable formulation of its CDF, numerical integration is used through Monte Carlo (MC) or Quasi-Monte Carlo (QMC) based methods, where this can refer to factorized or to non-factorisable multivariate distributions.
The method may be one wherein a copula is used as a multivariate cumulative distribution function.
The method may be one wherein to obtain a probability density function over the latent space, the corresponding characteristic function is transformed using a Fourier Transform to obtain the probability density function.
The method may be one wherein to evaluate joint probability distributions over the pixel space, an input of the latent space into the characteristic function space is transformed, and then the given/learned characteristic function is evaluated, and the output is converted back into the joint-spatial probability space.
The method may be one wherein to incorporate multimodality into entropy modelling, a mixture model is used as a prior distribution.
The method may be one wherein to incorporate multimodality into entropy modelling, a mixture model is used as a prior distribution, comprising a weighted sum of any base (parametric or non-parametric, factorized or non-factorisable multivariate) distribution as mixture components.
The method may be one wherein the (e.g. factorized) probability distribution is a non-parametric (e.g. factorized) probability distribution.
The method may be one wherein the non-parametric (e.g. factorized) probability distribution is a histogram model, or a kernel density estimation, or a learned (e.g. factorized) cumulative density function.
The method may be one wherein a prior distribution is imposed on the latent space, in which the prior distribution is an entropy model, which is optimized over its assigned parameter space to match its underlying distribution, which in turn lowers encoding computational operations.
The method may be one wherein the parameter space is sufficiently flexible to properly model the latent distribution.
The method may be one wherein encoding the quantized y latent using the third trained neural network, using the first computer system, to produce a z latent representation, includes using an invertible neural network, and wherein the second computer system processing the quantized z latent to produce the quantized y latent, includes using an inverse of the invertible neural network.
The method may be one wherein a hyperprior network of a compression pipeline is integrated with a normalising flow.
The method may be one wherein there is provided a modification to the architecture of normalising flows that introduces hyperprior networks in each factor-out block.
The method may be one wherein there is provided meta-compression, where the decoder weights are compressed with a normalising flow and sent along within the bitstreams.
The method may be one wherein encoding the input image using the first trained neural network includes using one or more univariate or multivariate Padé activation units.
The method may be one wherein using the second trained neural network to produce an output image from the quantized latent includes using one or more univariate or multivariate Padé activation units.
The method may be one wherein encoding the quantized y latent using the third trained neural network includes using one or more univariate or multivariate Padé activation units.
The method may be one wherein using the fourth trained neural network to obtain probability distribution parameters of each element of the quantized y latent includes using one or more univariate or multivariate Padé activation units.
The method may be one wherein steps (ii) to (xiii) are executed wholly in a frequency domain.
The method may be one wherein integral transforms to and from the frequency domain are used.
The method may be one wherein the integral transforms are Fourier Transforms, or Hartley Transforms, or Wavelet Transforms, or Chirplet Transforms, or Sine and Cosine Transforms, or Mellin Transforms, or Hankel Transforms, or Laplace Transforms.
The method may be one wherein spectral convolution is used for image compression.
The method may be one wherein spectral specific activation functions are used.
The method may be one wherein for downsampling, an input is divided into several blocks that are concatenated in a separate dimension; a convolution operation with a lxi kernel is then applied such that the number of channels is reduced by half; and wherein the upsampling follows a reverse and mirrored methodology.
The method may be one wherein for image decomposition, stacking is performed.
The method may be one wherein for image reconstruction, stitching is performed.
The method may be one wherein the first computer system is a server, e.g. a dedicated server, e.g a machine in the cloud with dedicated GPUs e.g Amazon Web Services, Microsoft Azure, etc, or any other cloud computing services.
The method may be one wherein the first computer system is a user device.
The method may be one wherein the user device is a laptop computer, desktop computer, a tablet computer or a smart phone.
The method may be one wherein the first trained neural network includes a library installed on the first computer system.
The method may be one wherein the first trained neural network is parametrized by one or several convolution matrices θ, or wherein the first trained neural network is parametrized by a set of bias parameters, non-linearity parameters, convolution kernel/matrix parameters.
The method may be one wherein the second computer system is a recipient device.
The method may be one wherein the recipient device is a laptop computer, desktop computer, a tablet computer, a smart TV or a smart phone.
The method may be one wherein the second trained neural network includes a library installed on the second computer system.
The method may be one wherein the second trained neural network is parametrized by one or several convolution matrices Ω, or wherein the first trained neural network is parametrized by a set of bias parameters, non-linearity parameters, convolution kernel/matrix parameters.
An advantage of the above is that for a fixed file size (“rate”), a reduced output image distortion may be obtained. An advantage of the above is that for a fixed output image distortion, a reduced file size (“rate”) may be obtained.
According to an eighth aspect of the invention, there is provided a system for lossy image or video compression, transmission and decoding, the system including a first computer system, a first trained neural network, a second computer system, a second trained neural network, a third trained neural network, a fourth trained neural network and a trained neural network identical to the fourth trained neural network, wherein:
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The system may be one wherein the system is configured to perform a method of any aspect of the seventh aspect of the invention.
According to a ninth aspect of the invention, there is provided a first computer system of any aspect of the eighth aspect of the invention.
According to a tenth aspect of the invention, there is provided a second computer system of any aspect of the eighth aspect of the invention.
According to an eleventh aspect of the invention, there is provided a computer implemented method of training a first neural network, a second neural network, a third neural network, and a fourth neural network, the neural networks being for use in lossy image or video compression, transmission and decoding, the method including the steps of:
An advantage of the invention is that, when using the trained first neural network, the trained second neural network, the trained third neural network and the trained fourth neural network, for a fixed file size (“rate”), a reduced output image distortion is obtained; and for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the loss function is evaluated as a weighted sum of differences between the output image and the input training image, and the estimated bits of the quantized image latents.
The method may be one wherein the steps of the method are performed by a computer system.
The method may be one wherein the loss function is a weighted sum of a rate and a distortion.
The method may be one wherein for differentiability, actual quantisation is replaced by noise quantisation.
The method may be one wherein the noise distribution is uniform, Gaussian or Laplacian distributed, or a Cauchy distribution, a Logistic distribution, a Student's t distribution, a Gumbel distribution, an Asymmetric Laplace distribution, a skew normal distribution, an exponential power distribution, a Johnson's SU distribution, a generalized normal distribution, or a generalized hyperbolic distribution, or any commonly known univariate or multivariate distribution.
The method may be one wherein encoding the input training image using the first neural network includes using one or more univariate or multivariate Padé activation units.
The method may be one wherein using the second neural network to produce an output image from the quantized y latent includes using one or more univariate or multivariate Padé activation units.
The method may be one wherein encoding the quantized y latent using the third neural network includes using one or more univariate or multivariate Padé activation units.
The method may be one wherein using the fourth neural network to obtain probability distribution parameters of each element of the quantized y latent includes using one or more univariate or multivariate Padé activation units.
The method may be one wherein when back-propagating the gradient of the loss function through the second neural network, through the fourth neural network, through the third neural network and through the first neural network, parameters of the one or more univariate or multivariate Padé activation units of the first neural network are updated, parameters of the one or more univariate or multivariate Padé activation units of the third neural network are updated, parameters of the one or more univariate or multivariate Padé activation units of the fourth neural network are updated, and parameters of the one or more univariate or multivariate Padé activation units of the second neural network are updated.
The method may be one wherein in step (ix), the parameters of the one or more univariate or multivariate Padé activation units of the first neural network are stored, the parameters of the one or more univariate or multivariate Padé activation units of the second neural network are stored, the parameters of the one or more univariate or multivariate Padé activation units of the third neural network are stored, and the parameters of the one or more univariate or multivariate Padé activation units of the fourth neural network are stored.
An advantage of the above is that, when using the trained first neural network, the trained second neural network, the trained third neural network and the trained fourth neural network, for a fixed file size (“rate”), a reduced output image distortion may be obtained; and for a fixed output image distortion, a reduced file size (“rate”) may be obtained.
According to a twelfth aspect of the invention, there is provided a computer program product for training a first neural network, a second neural network, a third neural network, and a fourth neural network, the neural networks being for use in lossy image or video compression, transmission and decoding, the computer program product executable on a processor to:
The computer program product may be one wherein the loss function is evaluated as a weighted sum of differences between the output image and the input training image, and the estimated bits of the quantized image latents.
The computer program product may be executable on the processor to perform a method of any aspect of the eleventh aspect of the invention.
According to a thirteenth aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein in step (viii) the output image is stored.
The method may be one wherein the segmentation algorithm is a classification-based segmentation algorithm, or an object-based segmentation algorithm, or a semantic segmentation algorithm, or an instance segmentation algorithm, or a clustering based segmentation algorithm, or a region-based segmentation algorithm, or an edge-detection segmentation algorithm, or a frequency based segmentation algorithm.
The method may be one wherein the segmentation algorithm is implemented using a neural network.
The method may be one wherein Just Noticeable Difference (JND) masks are provided as input into a compression pipeline.
The method may be one wherein JND masks are produced using Discrete Cosine Transform (DCT) and Inverse DCT on the image segments from the segmentation algorithm.
The method may be one wherein the segmentation algorithm is used in a bi-level fashion.
According to a fourteenth aspect of the invention, there is provided a computer implemented method of training a first neural network and a second neural network, the neural networks being for use in lossy image or video compression, transmission and decoding, the method including the steps of:
An advantage of the invention is that, when using the trained first neural network and the trained second neural network, for a fixed file size (“rate”), a reduced output image distortion is obtained; and for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the loss function is evaluated as a weighted sum of differences between the output image and the input training image, and the estimated bits of the quantized image latents.
The method may be one wherein the steps of the method are performed by a computer system.
The method may be one wherein the loss function is a sum of respective rate and respectively weighted respective distortion, over respective training image segments, of a plurality of training image segments.
The method may be one wherein a higher weight is given to training image segments which relate to human faces.
The method may be one wherein a higher weight is given to training image segments which relate to text.
The method may be one wherein the segmentation algorithm is implemented using a neural network.
The method may be one wherein the segmentation algorithm neural network is trained separately to the first neural network and to the second neural network.
The method may be one wherein the segmentation algorithm neural network is trained end-to-end with the first neural network and the second neural network.
The method may be one wherein gradients from the compression network do not affect the segmentation algorithm neural network training, and the segmentation network gradients do not affect the compression network gradients.
The method may be one wherein the training pipeline includes a plurality of Encoder; Decoder pairs, wherein each Encoder; Decoder pair produces patches with a particular loss function which determines the types of compression distortion each compression network produces.
The method may be one wherein the loss function is a sum of respective rate and respectively weighted respective distortion, over respective training image segments, of a plurality of training image colour segments.
The method may be one wherein an adversarial GAN loss is applied for high frequency regions, and an MSE is applied for low frequency areas.
The method may be one wherein a classifier trained to identify optimal distortion losses for image or video segments is used to train the first neural network and the second neural network.
The method may be one wherein the segmentation algorithm is trained in a bi-level fashion.
The method may be one wherein the segmentation algorithm is trained in a bi-level fashion to selectively apply losses for each segment during training of the first neural network and the second neural network.
An advantage of the above is that, when using the trained first neural network and the trained second neural network, for a fixed file size (“rate”), a reduced output image distortion may be obtained; and for a fixed output image distortion, a reduced file size (“rate”) may be obtained.
According to a fifteenth aspect of the invention, there is provided a classifier trained to identify optimal distortion losses for image or video segments, and usable in a computer implemented method of training a first neural network and a second neural network of any aspect of the fourteenth aspect of the invention.
According to a sixteenth aspect of the invention, there is provided a computer-implemented method for training a neural network to predict human preferences of compressed image segments for distortion types, the method including the steps of
According to a seventeenth aspect of the invention, there is provided a computer-implemented method for training neural networks for lossy image or video compression, trained with a segmentation loss with variable distortion based on estimated human preference, the method including the steps of
The method may be one wherein the loss function is evaluated as a weighted sum of differences between the output image and the input training image, and the estimated bits of the quantized image latents.
According to an eighteenth aspect of the invention, there is provided a computer implemented method of training a first neural network and a second neural network based on training images in which each respective training image includes human scored data relating to a perceived level of distortion in the respective training image as evaluated by a group of humans, the neural networks being for use in lossy image or video compression, transmission and decoding, the method including the steps of:
An advantage of the invention is that, when using the trained first neural network and the trained second neural network, for a fixed file size (“rate”), a reduced output image distortion is obtained; and for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the loss function is evaluated as a weighted sum of differences between the output image and the input training image, and the estimated bits of the quantized image latents.
The method may be one wherein the steps of the method are performed by a computer system.
The method may be one wherein at least one thousand training images are used.
The method may be one wherein the training images include a wide range of distortions.
The method may be one wherein the training images include mainly distortions introduced using AI-based compression encoder-decoder pipelines.
The method may be one wherein the human scored data is based on human labelled data.
The method may be one wherein in step (v) the loss function includes a component that represents the human visual system.
According to a nineteenth aspect of the invention, there is provided a computer-implemented method of learning a function from compression specific human labelled image data, the function suitable for use in a distortion function which is suitable for training an AI-based compression pipeline for images or video, the method including the steps of
The method may be one wherein other information (e.g. saliency masks), can be passed into the network along with the images too.
The method may be one wherein rate is used as a proxy to generate and automatically label data in order to pre-train the neural network.
The method may be one wherein ensemble methods are used to improve the robustness of the neural network.
The method may be one wherein multi-resolution methods are used to improve the performance of the neural network.
The method may be one wherein Bayesian methods are applied to the learning process.
The method may be one wherein a learned function is used to train a compression pipeline.
The method may be one wherein a learned function and MSE/PSNR are used to train a compression pipeline.
According to a twentieth aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that for a fixed file size (“rate”), a reduced output images distortion {circumflex over (x)}1, {circumflex over (x)}2 is obtained. An advantage of the invention is that for a fixed output images {circumflex over (x)}1, {circumflex over (x)}2 distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein in step (vii) the output pair of stereo images is stored.
The method may be one wherein ground-truth dependencies between x1, {circumflex over (x)}2 are used as additional input.
The method may be one wherein depth maps of x1, x2 are used as additional input.
The method may be one wherein optical flow data of x1, x2 are used as additional input.
According to a 21st aspect of the invention, there is provided a computer implemented method of training a first neural network and a second neural network, the neural networks being for use in lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that, when using the trained first neural network and the trained second neural network, for a fixed file size (“rate”), a reduced output images {circumflex over (x)}1, {circumflex over (x)}2 distortion is obtained; and for a fixed output images {circumflex over (x)}1, {circumflex over (x)}2 distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the loss function is evaluated as a weighted sum of differences between the output images and the input training images, and the estimated bits of the quantized image latents.
The method may be one wherein the steps of the method are performed by a computer system.
The method may be one wherein the loss function includes using a single image depth-map estimation of x1, x2, {circumflex over (x)}1, {circumflex over (x)}2 and then measuring the distortion between the depths maps of x1, {circumflex over (x)}1 and x2, {circumflex over (x)}2.
The method may be one wherein the loss function includes using a reprojection into the 3-d world using x1, x2, and one using {circumflex over (x)}1, {circumflex over (x)}2 and a loss measuring the difference of the resulting 3-d worlds.
The method may be one wherein the loss function includes using optical flow methods that establish correspondence between pixels in x1, x2 and {circumflex over (x)}1, {circumflex over (x)}2, and a loss to minimise these resulting flow-maps.
The method may be one wherein positional location information of the cameras/images and their absolute/relative configuration are encoded in the neural networks as a prior through the training process.
According to a 22nd aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that for a fixed file size (“rate”), a reduced N multi-view output images distortion is obtained. An advantage of the invention is that for a fixed N multi-view output images distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein in step (vii) the N multi-view output images are stored.
The method may be one wherein ground-truth dependencies between the N multi-view images are used as additional input.
The method may be one wherein depth maps of the N multi-view images are used as additional input.
The method may be one wherein optical flow data of the N multi-view images are used as additional input.
According to a 23rd aspect of the invention, there is provided a computer implemented method of training a first neural network and a second neural network, the neural networks being for use in lossy image or video compression, transmission and decoding, the method including the steps of:
An advantage of the invention is that, when using the trained first neural network and the trained second neural network, for a fixed file size (“rate”), a reduced N multi-view output images distortion is obtained; and for a fixed N multi-view output images distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the loss function is evaluated as a weighted sum of differences between the output images and the input training images, and the estimated bits of the quantized image latents.
The method may be one wherein the steps of the method are performed by a computer system.
The method may be one wherein the loss function includes using a single image depth-map estimation of the N multi-view input training images and the N multi-view output images and then measuring the distortion between the depth maps of the N multi-view input training images and the N multi-view output images.
The method may be one wherein the loss function includes using a reprojection into the 3-d world using N multi-view input training images and a reprojection into the 3-d world using N multi-view output images and a loss measuring the difference of the resulting 3-d worlds.
The method may be one wherein the loss function includes using optical flow methods that establish correspondence between pixels in N multi-view input training images and N multi-view output images and a loss to minimise these resulting flow-maps.
The method may be one wherein positional location information of the cameras/images and their absolute/relative configuration are encoded in the neural networks as a prior through the training process.
According to a 24th aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that for a fixed file size (“rate”), a reduced output satellite/space or medical image distortion is obtained. An advantage of the invention is that for a fixed output satellite/space or medical image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the output satellite/space, hyperspectral or medical image is stored.
According to a 25th aspect of the invention, there is provided a computer implemented method of training a first neural network and a second neural network, the neural networks being for use in lossy image or video compression, transmission and decoding, the method including the steps of:
An advantage of the invention is that, when using the trained first neural network and the trained second neural network, for a fixed file size (“rate”), a reduced output satellite/space or medical image distortion is obtained; and for a fixed output satellite/space or medical image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the loss function is evaluated as a weighted sum of differences between the output image and the input training image, and the estimated bits of the quantized image latents.
The method may be one wherein the steps of the method are performed by a computer system.
According to a 26th aspect of the invention, there is provided a computer implemented method of training a first neural network and a second neural network, the neural networks being for use in lossy image or video compression, transmission and decoding, the method including the steps of:
An advantage of the invention is that, when using the trained first neural network and the trained second neural network, for a fixed file size (“rate”), a reduced output image distortion is obtained; and for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the loss function is evaluated as a weighted sum of differences between the output image and the input training image, and the estimated bits of the quantized image latents.
The method may be one wherein the steps of the method are performed by a computer system.
The method may be one wherein the entropy loss includes moment matching.
According to a 27th aspect of the invention, there is provided a computer implemented method of training a first neural network and a second neural network, the method including the use of a discriminator neural network, the first neural network and the second neural network being for use in lossy image or video compression, transmission and decoding, the method including the steps of:
An advantage of the invention is that, when using the trained first neural network and the trained second neural network, for a fixed file size (“rate”), a reduced output image distortion is obtained; and for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the steps of the method are performed by a computer system.
The method may be one wherein the parameters of the trained discriminator neural network are stored.
According to a 28th aspect of the invention, there is provided a computer implemented method of training a first neural network and a second neural network, the neural networks being for use in lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that, when using the trained first neural network and the trained second neural network, for a fixed file size (“rate”), a reduced output image distortion is obtained; and for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the loss function is evaluated as a weighted sum of differences between the output image and the input training image, and the estimated bits of the quantized image latents.
The method may be one wherein the steps of the method are performed by a computer system.
According to a 29th aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein in step (vii) the output image is stored.
The method may be one wherein the routing network is trained using reinforcement learning.
The method may be one wherein the reinforcement learning includes continuous relaxation.
The method may be one wherein the reinforcement learning includes discrete k-best choices.
The method may be one wherein the training approach for optimising the loss/reward function for the routing module includes using a diversity loss.
The method may be one wherein the diversity loss is a temporal diversity loss, or a batch diversity loss.
According to a 30th aspect of the invention, there is provided a computer-implemented method, using a neural network architecture search (NAS) of determining one or multiple candidate architectures for a neural network for performing AI-based Image/Video Compression, the method including the steps of:
The method may be one wherein the method is applied to operator selection, or optimal neural cell creation, or optimal micro neural search, or optimal macro neural search.
The method may be one wherein a set of possible operators in the network is defined, wherein the problem of training the network is a discrete selection process and Reinforcement Learning tools are used to select a discrete operator per function at each position in the neural network.
The method may be one wherein the Reinforcement Learning treats this as an agent-world problem in which an agent has to choose the proper discrete operator, and the agent is training using a reward function.
The method may be one wherein Deep Reinforcement Learning, or Gaussian Processes, or Markov Decision Processes, or Dynamic Programming, or Monte Carlo Methods, or a Temporal Difference algorithm, are used.
The method may be one wherein a set of possible operators in the network is defined, wherein to train the network, Gradient-based NAS approaches are used by defining a specific operator as a linear (or non-linear) combination over all operators of the set of possible operators in the network; then, gradient descent is used to optimise the weight factors in the combination during training.
The method may be one wherein a loss is included to incentive the process to become less continuous and more discrete over time by encouraging one factor to dominate (e.g. GumbelMax with temperature annealing).
The method may be one wherein a neural architecture is determined for one or more of an Encoder, a Decoder, a Quantisation Function, an Entropy Model, an Autoregressive Module and a Loss Function.
The method may be one wherein the method is combined with auxiliary losses for AI-based Compression for compression-objective architecture training.
The method may be one wherein the auxiliary losses are runtime on specific hardware-architectures and/or devices, FLOP-count, memory-movement.
According to a 31st aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the finetuning loss measures one of, or a combination of: a rate of the modified quantized latent, or a distortion between the current decoder prediction of the output image and the input image, or a distortion between the current decoder prediction of the output image and a decoder prediction of the output image using the quantized latent from step (iii).
The method may be one wherein the loop in step (iv) ends when the modified quantized latent satisfies an optimization criterion.
The method may be one wherein in step (iv), the quantized latent is modified using a 1st-order optimization method, or using a 2nd-order optimization method, or using Monte-Carlo, Metropolis-Hastings, simulated annealing, or other greedy approaches.
According to a 32nd aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the finetuning loss measures one of, or a combination of a rate of the quantized latent, or a distortion between the current decoder prediction of the output image and the input image, or a distortion between the current decoder prediction of the output image and a decoder prediction of the output image using the quantized latent from step (iv).
The method may be one wherein the loop in step (iii) ends when the modified latent satisfies an optimization criterion.
The method may be one wherein in step (iii), the latent is modified using a 1st-order optimization method, or using a 2nd-order optimization method, or using Monte-Carlo, Metropolis-Hastings, simulated annealing, or other greedy approaches.
According to a 33rd aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the finetuning loss measures one of, or a combination of a rate of the quantized latent, or a distortion between the current decoder prediction of the output image and the input image, or a distortion between the current decoder prediction of the output image and a decoder prediction of the output image using the quantized latent from step (iv).
The method may be one wherein the loop in step (ii) ends when the modified input image satisfies an optimization criterion.
The method may be one wherein in step (ii), the input image is modified using a 1st-order optimization method, or using a 2nd-order optimization method, or using Monte-Carlo, Metropolis-Hastings, simulated annealing, or other greedy approaches.
According to a 34th aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the parameters are a discrete perturbation of the weights of the second trained neural network.
The method may be one wherein the weights of the second trained neural network are perturbed by a perturbation function that is a function of the parameters, using the parameters in the perturbation function.
According to a 35th aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of:
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein in step (iv), the binary mask is optimized using a ranking based method, or using a stochastic method, or using a sparsity regularization method.
According to a 36th aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the linear neural network is a purely linear neural network.
According to a 37th aspect of the invention, there is provided a computer-implemented method for lossy image or video compression, transmission and decoding, the method including the steps of:
An advantage of the invention is that for a fixed file size (“rate”), a reduced output image distortion is obtained. An advantage of the invention is that for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the linear neural network is a purely linear neural network.
According to a 38th aspect of the invention, there is provided a computer implemented method of training a first neural network, a second neural network, a third neural network, and a fourth neural network, the neural networks being for use in lossy image or video compression, transmission and decoding, the method including the steps of:
According to a 39th aspect of the invention, there is provided a computer implemented method of training a first neural network, a second neural network, a third neural network, and a fourth neural network, the neural networks being for use in lossy image or video compression, transmission and decoding, the method including the steps of:
An advantage of each of the above two inventions is that, when using the trained first neural network, the trained second neural network, the trained third neural network and the trained fourth neural network, for a fixed file size (“rate”), a reduced output image distortion is obtained; and for a fixed output image distortion, a reduced file size (“rate”) is obtained.
The method may be one wherein the loss function is evaluated as a weighted sum of differences between the output image and the input training image, and the estimated bits of the quantized image latents.
The method may be one wherein the steps of the method are performed by a computer system.
The method may be one wherein initially the units are stabilized by using a generalized convolution operation, and then after a first training the weights of the trained first neural network, the trained third neural network and the trained fourth neural network, are stored and frozen; and then in a second training process the generalized convolution operation of the units is relaxed, and the second neural network is trained, and its weights are then stored.
The method may be one wherein the second neural network is proxy trained with a regression operation.
The method may be one wherein the regression operation is linear regression, or Tikhonov regression.
The method may be one wherein initially the units are stabilized by using a generalized convolution operation or optimal convolution kernels given by linear regression and/or Tikhonov stabilized regression, and then after a first training the weights of the trained first neural network, the trained third neural network and the trained fourth neural network, are stored and frozen; and then in a second training process the generalized convolution operation is relaxed, and the second neural network is trained, and its weights are then stored.
The method may be one wherein in a first training period joint optimization is performed for a generalised convolution operation of the units, and a regression operation of the second neural network, with a weighted loss function, whose weighting is dynamically changed over the course of network training, and then the weights of the trained first neural network, the trained third neural network and the trained fourth neural network, are stored and frozen; and then in a second training process the generalized convolution operation of the units is relaxed, and the second neural network is trained, and its weights are then stored.
Aspects of the invention may be combined.
In the above methods and systems, an image may be a single image, or an image may be a video image, or images may be a set of video images, for example.
The above methods and systems may be applied in the video domain.
For each of the above methods, a related system may be provided.
For each of the above training methods, a related computer program product may be provided.
Aspects of the invention will now be described, by way of example(s), with reference to the following Figures, in which:
N to
N.
M is transformed through an encoder E, creating a latent y∈
n. The latent y is
quantized, becoming an integer-valued vector ŷ∈Zn. During training of the pipeline, a probability model on ŷ is used to compute estimate the rate R (the length of the bitstream). During use, the probability model is used by an arithmetic encoder & arithmetic decoder, which transform the quantized latent into a bitstream (and vice versa). On decode, the quantized latent is sent through a decoder D, returning a prediction {circumflex over (x)} approximating x.
We provide a high level overview of our artificial intelligence (AI)-based (e.g. image and/or video) compression technology.
In general, compression can be lossless, or lossy. In lossless compression, and in lossy compression, the file size is reduced. The file size is sometimes referred to as the “rate”.
But in lossy compression, it is possible to change what is input. The output image {circumflex over (x)} after reconstruction of a bitstream relating to a compressed image is not the same as the input image x. The fact that the output image {circumflex over (x)} may differ from the input image x is represented by the hat over the “x”. The difference between x and {circumflex over (x)} may be referred to as “distortion”, or “a difference in image quality”. Lossy compression may be characterized by the “output quality”, or “distortion”.
Although our pipeline may contain some lossless compression, overall the pipeline uses lossy compression.
Usually, as the rate goes up, the distortion goes down. A relation between these quantities for a given compression scheme is called the “rate-distortion equation”. For example, a goal in improving compression technology is to obtain reduced distortion, for a fixed size of a compressed file, which would provide an improved rate-distortion equation. For example, the distortion can be measured using the mean square error (MSE) between the pixels of x and {circumflex over (x)}, but there are many other ways of measuring distortion, as will be clear to the person skilled in the art. Known compression and decompression schemes include for example, JPEG, JPEG2000, AVC, HEVC, AVI.
Our approach includes using deep learning and AI to provide an improved compression and decompression scheme, or improved compression and decompression schemes.
In an example of an artificial intelligence (AI)-based compression process, an input image x is provided. There is provided a neural network characterized by a function E( . . . ) which encodes the input image x. This neural network E( . . . ) produces a latent representation, which we call y. The latent representation is quantized to provide ŷ, a quantized latent. The quantized latent goes to another neural network characterized by a function D( . . . ) which is a decoder. The decoder provides an output image, which we call {circumflex over (x)}. The quantized latent ŷ is entropy-encoded into a bitstream.
For example, the encoder is a library which is installed on a user device, e.g. laptop computer, desktop computer, smart phone. The encoder produces the y latent, which is quantized to ŷ, which is entropy encoded to provide the bitstream, and the bitstream is sent over the internet to a recipient device. The recipient device entropy decodes the bitstream to provide ŷ, and then uses the decoder which is a library installed on a recipient device (e.g. laptop computer, desktop computer, smart phone) to provide the output image {circumflex over (x)}.
E may be parametrized by a convolution matrix θ such that y=Eθ(x).
D may be parametrized by a convolution matrix Ω such that {circumflex over (x)}=DΩ(ŷ).
We need to find a way to learn the parameters θ and Ω of the neural networks.
The compression pipeline may be parametrized using a loss function L. In an example, we use back-propagation of gradient descent of the loss function, using the chain rule, to update the weight parameters of θ and Ω of the neural networks using the gradients ∂L/∂w.
The loss function is the rate-distortion trade off. The distortion function is (x, {circumflex over (x)}), which produces a value, which is the loss of the distortion
. The loss function can be used to back-propagate the gradient to train the neural networks.
So for example, we use an input image, we obtain a loss function, we perform a backwards propagation, and we train the neural networks. This is repeated for a training set of input images, until the pipeline is trained. The trained neural networks can then provide good quality output images.
An example image training set is the KODAK image set (e.g. at www.cs.albany.edu/˜xypan/research/snr/Kodak.html). An example image training set is the IMAX image set. An example image training set is the Imagenet dataset (e.g. at www.image-net.org/download). An example image training set is the CLIC Training Dataset P (“professional”) and M (“mobile”) (e.g. at http://challenge.compression.cc/tasks/).
In an example, the production of the bitstream from y is lossless compression.
Based on Shannon entropy in information theory, the minimum rate (which corresponds to the best possible lossless compression) is the sum from i=1 to N of (pŷ(ŷi)*log2(pŷ(ŷi))) bits, where pŷ is the probability of ŷ, for different discrete ŷ values ŷi, where ŷ={ŷ1, ŷ2 . . . ŷN}, where we know the probability distribution p. This is the minimum file size in bits for lossless compression of ŷ.
Various entropy encoding algorithms are known, e.g. range encoding/decoding, arithmetic encoding/decoding.
In an example, entropy coding EC uses ŷ and pŷ to provide the bitstream. In an example, entropy decoding ED takes the bitstream and pŷ and provides ŷ. This example coding/decoding process is lossless.
How can we get filesize in a differentiable way? We use Shannon entropy, or something similar to Shannon entropy. The expression for Shannon entropy is fully differentiable. A neural network needs a differentiable loss function. Shannon entropy is a theoretical minimum entropy value. The entropy coding we use may not reach the theoretical minimum value, but it is expected to reach close to the theoretical minimum value.
In the pipeline, the pipeline needs a loss that we can use for training, and the loss needs to resemble the rate-distortion trade off.
A loss which may be used for neural network training is Loss=+λ*R, where
is the distortion function, k is a weighting factor, and R is the rate loss. R is related to entropy. Both
and R are differentiable functions.
There are some problems concerning the rate equation.
The Shannon entropy H gives us some minimum file size as a function of ŷ and pŷ i.e. H(ŷ, pŷ). The problem is how can we know pŷ, the probability distribution of the input? Actually, we do not know pŷ. So we have to approximate pŷ. We use qŷ as an approximation to pŷ. Because we use qŷ instead of pŷ, we are instead evaluating a cross entropy rather than an entropy. The cross entropy CE(ŷ, qŷ) gives us the minimum filesize for ŷ given the probability distribution qŷ.
There is the relation
H(ŷ,pŷ)=CE(ŷ,qŷ)+KL(pŷ∥qŷ)
Where KL is the Kullback-Leibler divergence between pŷ and qŷ. The KL is zero, if pŷ and qŷ are identical.
In a perfect world we would use the Shannon entropy to train the rate equation, but that would mean knowing pŷ, which we do not know. We only know qŷ, which is an assumed distribution.
So to achieve small file compression sizes, we need qŷ to be as close as possible to pŷ. One category of our inventions relates to the qŷ we use.
In an example, we assume qŷ is a factorized parametric distribution.
One of our innovations is to make the assumptions about qv more flexible. This can enable qŷ to better approximate pŷ, thereby reducing the compressed filesize.
As an example, consider that pŷ is a multivariate normal distribution, with a mean μ vector and a covariant matrix Σ. Σ has the size N×N, where N is the number of pixels in the latent space. Assuming ŷ with dimensions 1×12×512×512 (relating to images with e.g. 512×512 pixels), then Σ has the size 2.5 million squared, which is about 5 trillion, so therefore there are 5 trillion parameters in E we need to estimate. This is not computationally feasible. So, usually, assuming a multivariate normal distribution is not computationally feasible.
Let us consider pŷ, which as we have argued is too complex to be known exactly. This joint probability density function p(ŷ) can be represented as a conditional probability function, as the second line of the equation below expresses.
Very often p(ŷ) is approximated by a factorized probability density function
p(ŷ1)*p(ŷ2)*p(ŷ3)* . . . p(ŷN)
The factorized probability density function is relatively easy to calculate computationally. One of our approaches is to start with a q which is a factorized probability density function, and then we weaken this condition so as to approach the conditional probability function, or the joint probability density function p(ŷ), to obtain smaller compressed filzesizes. This is one of the class of innovations that we have.
Distortion functions (x, {circumflex over (x)}), which correlate well with the human vision system, are hard to identify. There exist many candidate distortion functions, but typically these do not correlate well with the human vision system, when considering a wide variety of possible distortions.
We want humans who view picture or video content on their devices, to have a pleasing visual experience when viewing this content, for the smallest possible file size transmitted to the devices. So we have focused on providing improved distortion functions, which correlate better with the human vision system. Modern distortion functions very often contain a neural network, which transforms the input and the output into a perceptional space, before comparing the input and the output. The neural network can be a generative adversarial network (GAN) which performs some hallucination. There can also be some stabilization. It turns out it seems that humans evaluate image quality over density functions. We try to get p({circumflex over (x)}) to match p(x), for example using a generative method eg. a GAN.
Hallucinating is providing fine detail in an image, which can be generated for the viewer, where all the fine, higher spatial frequencies, detail does not need to be accurately transmitted, but some of the fine detail can be generated at the receiver end, given suitable cues for generating the fine details, where the cues are sent from the transmitter.
How should the neural networks E( . . . ), D( . . . ) look like? What is the architecture optimization for these neural networks? How do we optimize performance of these neural networks, where performance relates to filesize, distortion and runtime performance in real time? There are trade offs between these goals. So for example if we increase the size of the neural networks, then distortion can be reduced, and/or filesize can be reduced, but then runtime performance goes down, because bigger neural networks require more computational resources. Architecture optimization for these neural networks makes computationally demanding neural networks run faster.
We have provided innovation with respect to the quantization function Q. The problem with a standard quantization function is that it has zero gradient, and this impedes training in a neural network environment, which relies on the back propagation of gradient descent of the loss function. Therefore we have provided custom gradient functions, which allow the propagation of gradients, to permit neural network training.
We can perform post-processing which affects the output image. We can include in the bitstream additional information. This additional information can be information about the convolution matrix Ω, where D is parametrized by the convolution matrix Ω. The additional information about the convolution matrix Ω can be image-specific. An existing convolution matrix can be updated with the additional information about the convolution matrix Ω, and decoding is then performed using the updated convolution matrix.
Another option is to fine tune the y, by using additional information about E. The additional information about E can be image-specific.
The entropy decoding process should have access to the same probability distribution, if any, that was used in the entropy encoding process. It is possible that there exists some probability distribution for the entropy encoding process that is also used for the entropy decoding process. This probability distribution may be one to which all users are given access; this probability distribution may be included in a compression library; this probability distribution may be included in a decompression library. It is also possible that the entropy encoding process produces a probability distribution that is also used for the entropy decoding process, where the entropy decoding process is given access to the produced probability distribution. The entropy decoding process may be given access to the produced probability distribution by the inclusion of parameters characterizing the produced probability distribution in the bitstream. The produced probability distribution may be an image-specific probability distribution.
In an example of a layer in an encoder neural network, the layer includes a convolution, a bias and an activation function. In an example, four such layers are used.
In an example, we assume that qŷ is a factorized normal distribution, where y={y1, y2 . . . yN}, and ŷ={ŷ1, ŷ2 . . . ŷN}. We assume each ŷi(i=1 to N) follows a normal distribution N e.g. with a mean p of zero and a standard deviation σ of 1. We can define ŷ=Int(y−μ)+μ, where Into is integer rounding.
The rate loss in the quantized latent space comes from, summing (Σ) from i=1 to N,
Rate=(Σ log2(qŷ(ŷi)))/N=(ΣN(ŷi|μ=0,σ=1))/N
The output image {circumflex over (x)} can be sent to a discriminator network, e.g. a GAN network, to provide scores, and the scores are combined to provide a distortion loss.
We want to make the qŷ flexible so we can model the pŷ better, and close the gap between the Shannon entropy and the cross entropy. We make the qŷ more flexible by using meta information. We have another neural network on our y latent space which is a hyper encoder. We have another latent space called z, which is quantized to {circumflex over (z)}. Then we decode the z latent space into distribution parameters such as μ and α. These distribution parameters are used in the rate equation.
Now in the more flexible distribution, the rate loss is, summing (Σ) from i=1 to N,
Rate=(ΣN(ŷi|μi,σi))/N
So we make the q more flexible, but the cost is that we must send meta information. In this system, we have
bitstreamŷ=EC(ŷ,qŷ(μ,σ))
ŷ=ED(bitstreamŷ,qŷ(μ,σ))
Here the z latent gets its own bitstream{circumflex over (z)} which is sent with bitstreamŷ. The decoder then decodes bitstream{circumflex over (z)} first, then executes the hyper decoder, to obtain the distribution parameters (μ, σ), then the distribution parameters (μ, σ) are used with bitstreamŷ to decode the ŷ, which are then executed by the decoder to get the output image {circumflex over (x)}.
Although we now have to send bitstream{circumflex over (z)}, the effect of bitstream{circumflex over (z)} is that it makes bitstreamŷ smaller, and the total of the new bitstreamŷ and bitstream{circumflex over (z)} is smaller than bitstreamŷ without the use of the hyper encoder. This is a powerful method called hyperprior, and it makes the entropy model more flexible by sending meta information. The loss equation becomes
Loss=(x,{circumflex over (x)})+λ1*Ry+λ2*Rz
It is possible further to use a hyper hyper encoder for z, optionally and so on recursively, in more sophisticated approaches.
The entropy decoding process of the quantized z latent should have access to the same probability distribution, if any, that was used in the entropy encoding process of the quantized z latent. It is possible that there exists some probability distribution for the entropy encoding process of the quantized z latent that is also used for the entropy decoding process of the quantized z latent. This probability distribution may be one to which all users are given access; this probability distribution may be included in a compression library; this probability distribution may be included in a decompression library. It is also possible that the entropy encoding process of the quantized z latent produces a probability distribution that is also used for the entropy decoding process of the quantized z latent, where the entropy decoding process of the quantized z latent is given access to the produced probability distribution. The entropy decoding process of the quantized z latent may be given access to the produced probability distribution by the inclusion of parameters characterizing the produced probability distribution in the bitstream. The produced probability distribution may be an image-specific probability distribution.
In a more sophisticated approach, the distortion function (x, {circumflex over (x)}) has multiple contributions. The discriminator networks produce a generative loss LGEN. For example a Visual Geometry Group (VGG) network may be used to process x to provide m, and to process {circumflex over (x)} to provide {circumflex over (m)}, then a mean squared error (MSE) is provided using m and {circumflex over (m)} as inputs, to provide a perceptual loss. The MSE using x and {circumflex over (x)} as inputs, can also be calculated. The loss equation becomes
Loss=λ1*Ry+λ2*Rz+λ3*MSE(x,{circumflex over (x)})+λ4*LGEN+λ5*VGG(x,{circumflex over (x)}),
where the first two terms in the summation are the rate loss, and where the final three terms in the summation are the distortion loss (x, {circumflex over (x)}). Sometimes there can be additional regularization losses, which are there as part of making training stable.
Notes re HyperPrior and HyperHyperPrior
Regarding a system or method not including a hyperprior, if we have a y latent without a HyperPrior (i.e. without a third and a fourth network), the distribution over the y latent used for entropy coding is not thereby made flexible. The HyperPrior makes the distribution over the y latent more flexible and thus reduces entropy/filesize. Why? Because we can send y-distribution parameters via the HyperPrior. If we use a HyperPrior, we obtain a new, z, latent. This z latent has the same problem as the “old y latent” when there was no hyperprior, in that it has no flexible distribution. However, as the dimensionality re z usually is smaller than re y, the issue is less severe.
We can apply the concept of the HyperPrior recursively and use a HyperHyperPrior on the z latent space of the HyperPrior. If we have a z latent without a HyperHyperPrior (i.e. without a fifth and a sixth network), the distribution over the z latent used for entropy coding is not thereby made flexible. The HyperHyperPrior makes the distribution over the z latent more flexible and thus reduces entropy/filesize. Why? Because we can send z-distribution parameters via the HyperHyperPrior. If we use the HyperHyperPrior, we end up with a new w latent. This w latent has the same problem as the “old z latent” when there was no hyperhyperprior, in that it has no flexible distribution. However, as the dimensionality re w usually is smaller than re z, the issue is less severe. An example is shown in
The above-mentioned concept can be applied recursively. We can have as many HyperPriors as desired, for instance: a HyperHyperPrior, a HyperHyperHyperPrior, a HyperHyperHyperHyperPrior, and so on.
Notes re Training
Regarding seeding the neural networks for training, all the neural network parameters can be randomized with standard methods (such as Xavier Initialization). Typically, we find that satisfactory results are obtained with sufficiently small learning rates.
Note
It is to be understood that the arrangements referenced herein are only illustrative of the application for the principles of the present inventions. Numerous modifications and alternative arrangements can be devised without departing from the spirit and scope of the present inventions. While the present inventions are shown in the drawings and fully described with particularity and detail in connection with what is presently deemed to be the most practical and preferred examples of the inventions, it will be apparent to those of ordinary skill in the art that numerous modifications can be made without departing from the principles and concepts of the inventions as set forth herein.
Within the domain of learnt image and video compression, progress may be essentially measured jointly by two orthogonal metrics: perceptual quality and the compression factor of images. Perceptual quality can be hard to measure; a function for it may be completely intractable. Nevertheless, it is well known that the sensitivity of the human visual system (HVS) to different attributes in images, such as textures, colours and various objects, are different-humans are more likely to be able to identify an alteration performed to a human face compared to a patch of grass. By producing segments of images to which the HVS is more or less sensitive we can therefore improve the overall perceptual experience of the compressed media by optimising the learnt compression pipeline to follow heuristics from the HVS. We provide a modifications to the learnt compression pipeline that utilises a generic family of segmentation based approaches to allow the optimisation of the learnt compression network to more closely follow the heuristics of the HVS, achieving better perceptual quality at the same or at a higher compression factor.
Modern machine learning algorithms are optimised using a method called stochastic gradient descent. This method allows us to update the parameters of our model to a user-specific, desired goal. The goal is controlled by defining a loss-function that the network uses for backpropagation. Every parameter in the network is updated such that the loss is decreased as the network trains. In typical compression networks the same loss is applied to the entire image, see Equation (1.1).
The loss function within learnt compression can in its simplest form be considered to be composed of two different terms: one term that controls the distortion of the compressed image or video, D, and another term that controls the size of the compressed media (rate) R which is typically measured as the number of bits required per pixel (bpp). An uncompressed image requires 24 bpp, most compressed images are below 0.5 bpp. The λ parameter controls the trade-off between the size of the image and the compression distortions. For example, in the extreme case that λ=106, the value of R can become very large (lossless compression), since λD will be very large. In the other extreme λ=10−6, the network will be forced to learn such that R becomes very small (since AD is already minuscule).=R+ΔD (1.1)
In the standard learnt compression pipelines for image and video, Equation (1.1) is applied to train the network: is minimised. However a key question in the equation above is how the distortion D is estimated. Almost universally, the distortion of the media D, is computed in the same way across the entire image or video. Similarly, the constraint on the size R is computed the same for the entire image. Intuitively, it should be clear that some parts of the image should be assigned more bits, and some regions of the image should be prioritised in terms of image quality.
The reason for this intuition comes from the human visual system (HVS). It has been shown that humans are more susceptible to image degradations (such as compression artifacts) introduced in parts of the image that are more visually sensitive to the HVS. For example, the degradation of human faces or low frequency areas in the image are more noticeable to the HVS, and is therefore more likely to reduce the perceptual quality of the media. A mechanism of improving perceptual quality is thus to optimise parts of the image with different losses. To do this, we provide a generic modifications to the learnt compression pipeline powered by image segmentation operations, to compute dynamic losses optimised for the HVS.
In this section, a short introduction to the meaning of image segmentation within the field of computer vision is provided.
In the field of computer vision, image segmentation is a process that involves dividing a visual input into different segments based on some type of image analysis. Segments represent object or parts of objects, and comprise sets or groups of pixels. Image segmentation is a method of grouping pixels of the input into larger components. In the computer vision there are many different methods in which the segmentation may be performed to generate a grouping of pixels. A non-exhaustive list is provided below to provide examples:
The segmented images are typically produced by a neural network. However for the pipeline presented here, the segmentation operator can be completely generic.
An example of a generic pipeline is shown in
The loss function shown above in Equation (1.1) can therefore be modified as follows:
where n refers to the number of segments in the image, Ri is the rate for a particular segment, Di is the distortion for a particular segment, λi is the Lagrange multiplier and ci a constant, for segment i. This means that for each segment i in the image a different method of computing rate R and distortion D can be applied. For example, the distortion metric for texts may utilise an MSE loss, whereas for faces it utilises a mixture of perceptual and adversarial losses.
Furthermore, the computation of Ri means that each segment can have a variable rate. For example, assigning more bits to regions with higher sensitivity for the HVS, such as the faces and texts, or any other salient region in the image, will improve perceptual quality without increasing the total number of bits required for the compressed media.
This generic pipeline has been exemplified with 4 different segmentation approaches in the next section, however it extends to all types of segmentation, in addition to the 4 examples provided, such as clustering based segmentation, region-based segmentation, edge-detection segmentation, frequency based segmentation, any type of neural network powered segmentation approach, etc.
The segmentation module in
C
H×W×C
C({circumflex over (x)}, x, {circumflex over (x)}s))
C
s
H×W×C
H×W×C
H×W×C
s({circumflex over (x)}s, ys))
C({circumflex over (x)}, x, {circumflex over (x)}s))
In
It is well known that the HVS is more sensitive to changes in low frequency regions, such as uniform areas, compared to changes in high frequency regions such as in patches of grass. In general, for most images the majority of high frequencies can be removed within any noticeable difference in the image. Based on this intuition, it is therefore possible to create Just Noticeable Difference (JND) masks, based on segments of frequencies in the image, that indicate which parts of the image are most likely to be noticed by the HVS if distorted. One method by which the masks may be computed is using Algorithm 1.3.
Based on Algorithm 1.3, an example method of producing JND masks, is to use the Discrete Cosine Transform (DCT) and Inverse DCT on the segments from the segmentation operator. The JND masks may then be provided as input into the compression pipeline, for example, as shown in
N →
N
H×W×C
A different type of segmentation approach that more directly targets the HVS is to utilise a number of different learnt compression pipelines with distinctly different distortion metrics applied on the same segmentations of the images. Once trained, human raters are asked in a 2AFC selection procedure to indicate which patch from the trained compression pipelines produces the perceptually most pleasing image patch. For example, if there are 4 distortion metrics {d0, d1, d2, d3}, there will be 4 predicted patches, {{circumflex over (x)}0, {circumflex over (x)}1, {circumflex over (x)}2, {circumflex over (x)}3}, one from each metric trained with the different distortion losses {L0, L1, L2, L3}, as shown in
The image segmentation approaches discussed above segments pixels across the channels within the RGB colour space. However an alternative colour-space representation is known as YCbCr, where Y represents the luma component of the image; CbCr the chroma information of the image. Given a particular distortion metric that only operates on a certain portion of the colour space, a natural segmentation of the total distortion loss of the network is then an expectation of some number of distortion metrics across the colour space, where each component of the colour space may have a different distortion metric. That is, for example, some particular set of distortion metrics may operate on the luma component, whereas some other set may operate on the chroma part. The loss operating on each component has been optimized for the colour space in which it operates (or may not even be applicable outside the given space).
That is, the loss function may be re-written as below
where i is now an index into the colour space, where Ri, λi and Di refer to colour-space specific metrics.
The idea of colour-space segmentation is not limited to RGB and YCbCr, and is easily applied to any colour-space, such as CMYK, scRGB, CIE RGB, YPbPr, xvYCC, HSV, HSB, HSL, HLS, HSI, CIEXYZ, sRGB, ICtCp, CIELUV, CIEUVW, CIELAB, etc, as shown in
Accurate modelling of the true latent distribution is instrumental for minimising the rate term in a dual rate-distortion optimisation objective. A prior distribution imposed on the latent space, the entropy model, optimises over its assigned parameter space to match its underlying distribution, which in turn lowers encoding costs. Thus, the parameter space must be sufficiently flexible in order to properly model the latent distribution; here we provide a range of various methods to encourage flexibility in the entropy model.
In AI-based data compression, an autoencoder is a class of neural network whose parameters are tuned, in training, primarily to perform the following two tasks jointly:
Here we assume a lossy compression pipeline, however it should be noted that many concepts presented here are also applicable in lossless compression. The aforementioned tasks form the framework of a joint optimisation problem of two loss terms commonly found in compression problems, namely the minimisation of metrics representing rate, R(·), and distortion, D(·), respectively. The rate-distortion minimisation objective can mathematically be expressed in form of a weighted sum denoted by (·)
where x is the input data, θ is the network parameters and λ is a weighting factor that controls the rate-distortion balance. The rate loss is directly controlled by the ability of the network to accurately model the distribution of the latent representations of the input data, which brings forward the notion of entropy modelling which shall be outlined and justified in detail. In theory, the more accurately the entropy model matches the true latent distribution, the lower the rate term is. Note that the distortion term is also influenced indirectly as a result from the joint rate-distortion minimisation objective. However, for the sake of clarity, we will largely ignore the distortion term or any consequential impact of it from minimising the rate here.
The focus here is to
Below follows a detailed section on mathematical preliminaries that will act as a helpful guide. One common conventionality adopted is that the array data format of a quantity (scalars, vectors, matrices, etc.) is independent of the data itself. This means that if a quantity is fundamentally one-dimensional, such as a vector x of length N, then it can either be represented directly as a vector format x∈N or as an array (or tensor) format x∈
H×W×3 (where N=H×W×3). In other words, no matter how we organise x into different data structures, the fundamental variables contained within a particular instance of x are not mutated.
The following is a list of how various quantity types encountered within the text body are conventionally denoted:
The rest of the symbols commonly encountered relate to functions, operations or mappings, which follows the standardised form as detailed below:
The standard convention for index subscripts is the following: to index an individual element in a vector x, the subscript i is used for the column index (e.g. xi). To index an individual element in a matrix Σ, the subscripts i, j and lowercase letters are used for the row and column index, respectively (e.g. σi,j). Quantities with bracketed superscripts are associated with additional partitioning or groupings of vectors/matrices, such as latent space partitioning (often with index [b]) or base distribution component of a mixture model (often with index [k]). For example, indexing can look like y[b], ∀b∈{1, . . . , B} and μ[k], ∀k∈{1, . . . , K}.
This section serves to inform about the fundamentals of rate minimisation through entropy modelling of the latent distribution. We describe the various components in the network that this affects, why these components are necessary and the theory that underpins them. Demonstrative examples are also included as a guide.
The autoencoder for AI-based data compression, in a basic form, includes four main components:
The encoder transforms an N-dimensional input vector x to an M-dimensional latent vector y, hence the encoder transforms a data instance from input space to latent space (also called “bottleneck”) ƒenc: N→
M. M is generally smaller than N, although this is by no means necessary. The latent vector, or just the latents, acts as the transform coefficient which carries the source signal of the input data. Hence, the information in the data transmission emanates from the latent space.
As produced by the encoder, the latents generally comprise continuous floating point values. However, the transmission of floating point values directly is costly, since the idea of entropy coding does not lend itself well to continuous data. Hence, one technique is to discretise the latent space in a process called quantisation Q: M→
QM (where
QM denotes the quantised M-dimensional vector space,
QM⊂
M). During quantisation, latents are clustered into predetermined bins according to their value, and mapped to a fixed centroid of that bin. One way of doing this is by rounding the latents to the nearest integer value. The overall effect is that the set of possible values for the latents is reduced significantly which allows for shorter descriptors, but this also curbs expressiveness due to the irrecoverable information loss. We normally denote quantities that have undergone quantisation with a hat symbol, such as ŷ.
Once the latents are discretised, we can encode them into a bitstream. This process is called entropy coding which is a lossless encoding scheme; examples include arithmetic/range coding and Huffman coding. The entropy code comprises a codebook which uniquely maps each symbol (such as an integer value) to a binary codeword (comprised by bits, so 0s and 1s). These codewords are uniquely decodable, which essentially means in a continuous stream of binary codewords, there exists no ambiguity of the interpretation of each codeword. The optimal entropy code has a codebook that produces the shortest bitstream. This can be done by assigning the shorter codewords to the symbols with high probability, in the sense that we would transmit those symbols more times than less probable symbols. However, this requires knowing the probability distribution in advance.
This is where the entropy model comes in. It defines a prior probability distribution over the quantised latent space Pŷ(ŷ; ϕ), parametrised by the entropy parameters ϕ. The prior aims to model the true quantised latent distribution, also called the marginal distribution m(ŷ) which arises from what actually gets outputted by the encoder and quantisation steps, as closely as possible. The marginal is an unknown distribution; hence, the codebook in our entropy code is determined by the prior distribution whose parameters we can optimise for during training. The closer the prior models the marginal, the more optimal our entropy code mapping becomes which results in lower bitrates.
It is assumed that the codebook defined by the entropy model exists on both sides of the transmission channel. Under this condition, the transmitter can map a quantised latent vector into a bitstream, send it across the channel. The receiver can then decode the quantised latent vector from the bitstream losslessly, pass it through the decoder which transforms it into an approximation of the input vector {circumflex over (x)}, QM→
N.
What has been presented thus far is how a typical compression pipeline would work in practical application. However, during gradient descent-based training, we must ensure differentiability throughout the entire autoencoder in order for the loss gradients to backpropagate and update the network parameters. However, essential steps such as quantisation and entropy coding are usually non-differentiable and break the flow of gradient information during backpropagation. Therefore, an autoencoder often trains with proxy operations that mimic the prohibited operations whilst ensuring differentiability throughout the network. Specifically, we need to estimate the rate given our entropy model and simulate the effects of quantisation in a differentiable manner. Once the network has finished training, non-differentiable operations can be permitted for inference and real-life application.
Hence, we need to pay attention to the different “modes” of the network when it processes data; the particular “mode” of the network governs how certain operations behave within the network (see Table 2.1):
Information theory states that given a PMF MX(x) describing the probability distribution of the discrete random variable X, the shortest average message length that unambiguously relays information about a sample xi drawn from it is equal to the Shannon entropy of that distribution. The Shannon entropy is defined as
where b denotes the base of the logarithm. If b=2, the unit of this entropy is bits. This is where the notion of the optimal codebook in entropy coding, as well as the term entropy modelling, are derived from.
However, suppose we do not know the exact probability distribution of states (MX is unknown), but build our codebook with another known distribution Px(x), the average message length that unambiguously relays information about a sample xi drawn from MX is then equal to the cross-entropy of the distribution PX over MX:
The cross-entropy can be rephrased in terms of the Kullback-Leibler (KL) divergence, which is always nonnegative and can be interpreted as measuring how different two distributions are to one and another:
H(MX,PX)≡H(MX)+DKL(MX∥PX) (2.4)
From this, it is evident that the cross-entropy term is lower bounded by the Shannon entropy. If the cross-entropy reduces as a consequence of configuring PX, the KL divergence reduces commensurately, implying that PX is becoming more similar to MX. It is now clear what the motivation for learning a prior distribution Pŷ for the quantised latent space that ideally should match the unknown marginal distribution mŷ. The cross-entropy of Pŷ over mŷ acts as a theoretical measure for the achieved bitrate if we were to perform entropy coding with it, which is differentiable since it only depends on a logarithm operator and expectation operation! Hence, we can define our rate loss R by estimating the cross-entropy of the prior over the marginal:
R=H(mŷ,Pŷ)=ŷ˜m[log2Pŷ(ŷ)] (2.5)
Note that quantisation, whilst closely related to the entropy model, is a significant separate topic of its own. However, since quantisation influences certain aspects of entropy modelling, it is therefore important to briefly discuss the topic here. Specifically, they relate to
So far, we have only considered discrete probability distributions as entropy models. This is due to quantisation, which discretises the continuous vector space for the (raw) latents y. However, discrete distributions do not lend themselves well to gradient-based approaches due to their discontinuities. It is also possible to pick a continuous distribution as a prior, with the PDF py(y; ϕ) that is parametrised by ϕ, on the latent space. We can simply account for quantisation in the entropy model by evaluating probability masses over py by integrating over a zero-centred integration region Ω for each quantisation interval. For example, for a single variable ŷi (so in 1-D), the PMF can be defined as
Pŷi(ŷi)=∫ŷi+aiŷi+biPyi(y)dy,∀ŷi∈Q (2.6)
where the integration region per quantisation bin is defined by its bounds Ωi=[ai, bi]. In other words, the effect of quantisation on entropy modelling is that probability masses for each quantised state must be computed (for example, see
PyQ (2.7)
Example: Suppose the entropy model py(μi, σi). Its CDF is denoted by
where Φ(·) is the CDF of the standard normal distribution. Then, assuming regular integer-sized quantisation bins
we calculate the probability masses as follows:
The second point becomes slightly more involved. Here we will not discuss differentiable quantisation in more detail than necessary. The main discussion point revolve around perturbing y∈M with additive noise to simulate the effects of quantisation (there exist other differentiable quantisation methods which are known to the skilled person). Certain quantisation operations can be seen as having similar effects. Hence, when differentiability is imperative, we can substitute actual quantisation with noise quantisation {tilde over (Q)}: y
{tilde over (y)}
{tilde over (Q)}(y)={tilde over (y)}=y+∈Q (2.8)
where ∈Q is drawn from any random noise source distribution Θ, ideally similarly bounded as the perturbation emerging from actual quantisation though this is not a necessity. The simulated noise source Θ could theoretically have different distribution characteristics from the true quantisation perturbation source (it could for instance be Uniform, Gaussian or Laplacian distributed).
One key feature with this type of quantisation simulation is the effect it has on the (continuous) prior distribution. Unlike actual quantisation, {tilde over (Q)} maps a vector from M to
M, and not to the centroid of some quantisation bin. If we select the random noise source to be a uniform distribution with a width equal to the quantisation interval, the distribution of {tilde over (y)}, p{tilde over (y)}({tilde over (y)}) becomes a continuous relaxation of the probability mass formulation (Equation (2.6)). This can be understood by viewing the prior distribution as being convolved with the uniform distribution, which acts as a box-car smoothing filter (see rectangular box in
p{tilde over (y)}({tilde over (y)})=(py*P∈
Example: Suppose that the actual quantisation operation is rounding to the nearest integer, Q (y)=└y┐. This can be seen as adding a half-integer bounded noise vector
Hence, we can simulate the quantisation perturbation in training by adding a uniformly distributed random noise vector ∈Q, each element sampled from
This results in the continuously relaxed probability model
The true latent distribution of y∈M can be expressed, without loss of generality, as a joint (multivariate) probability distribution with conditionally dependent variables
p(y)≡p(Y1,Y2, . . . ,YM) (2.10)
which models the probability density over all sets of realisations of y. Therefore, it captures how each variable is distributed independently of the others as well as any intervariable dependencies between pairs of variables. However, since M is often very large, modelling intervariable dependencies between M variables would require enormous computational resources.
Another way to phrase a joint distribution is to evaluate the product of conditional distributions of each individual variable, given all previous variables:
P(Y1,y2, . . . ,yM)=P(Y1)·P(y2|y1)·P(y3|Y1,y2)· . . . ·P(YM|Y1, . . . ,yM-1) (2.11)
We can model each conditional distribution p(yi|y1, . . . , yi−1) using a so-called conditional or context model ƒcontext (·), which is a function mapping that takes in the previous variables and outputs the entropy parameters of the current variable: ϕi=ƒcontext({y1, . . . , yi−1}). In practice, ϕi would be evaluated one by one, which implies a serial encoding and decoding process. Assuming ideal parametrisation of the conditional distributions (which is rarely the case), we would be able to model the joint distribution perfectly. Unfortunately, serial encoding and decoding processes are very slow to execute, especially over a large number of dimensions.
Thus, in order to ensure realistic runtime of the operations, it is possible to ignore the conditioned variables, and model the latent distribution as a product of independent, univariate distributions
p(Y)=p(y1)·p(y2)·p(y3)· . . . ·P(YM) (2.12)
where each distribution p(yi) can be parametrised by entropy parameters ϕi. This type of entropy model is called factorised prior, since we can evaluate the factors (probability masses) for each variable individually (i.e. the joint is factorisable). The entropy parameters ϕ can be included with the network parameters that are optimised over during training, for which the termfully factorised is often used. The distribution type may be either parametric or non-parametric, with potentially multiple peaks and modes. See
AI-based data compression architectures may contain an additional autoencoder module, termed a hypernetwork. A hyperencoder henc(·) compresses metainformation in the form of hyperlatents z analogously to the main latents. Then, after quantisation, the hyperlatents are transformed through a hyperdecoder hdec(·) into instance-specific entropy parameters ϕ(see
However, a factorised prior ignores the notion of any dependency structure. This means that if the true latent distribution does have intervariable dependencies, a factorised prior would not be able to model these; the equal sign in Equation (2.12) would become an approximation sign. Thus, by Equation (2.4), it would never attain optimal compression performance (see
We have been very prolific in pushing the frontiers of entropy modelling by rigorous development of theory and experimental tests. This section introduces a range of innovations in this field. Outlined innovations are segmented in different categories, which are accordingly presented in the upcoming subsections. The categories are:
Some entropy models in AI-based data compression pipelines include factorised priors py
Therefore, we incorporate more flexibility in entropy modelling by using parametric distributions as factorised prior. We achieve this by employing distributions with many degrees of freedom in the parametrisation, including shape, asymmetry and skewness. Note that the innovation is formulated irrespective of the method with which the parameters ϕ are produced; these may be learned directly as fixed parameters (fully factorised prior), predicted by a hypernetwork (hyperprior) or by a context model (conditional model).
An example of parametric distribution families for factorised entropy modelling covered by this innovation, with the respective parametrisations for each distribution, can be seen in
Example: The exponential power distribution is a parametric family of continuous symmetric distributions. Apart from a location parameter μ and scale parameter α, it also includes a shape parameter, β>0. The PDF py (y), in the 1-D case, can be expressed as
where Γ(·) denotes the gamma function. The shape parameter can be modulated so as to attain probability densities from the normal (α=2), Laplace (β=1) and uniform (β=∞) distribution families, and a continuum of densities for any arbitrary ŷ value.
We have put a lot of effort into extending this approach to allow the quantised latent space to be modelled using discrete parametric probability distributions, as opposed to continuous probability distributions. Amongst others, we have tested and modified the following distributions to work in an AI-based data compression pipeline:
The entropy parameters in a compression pipeline define a probability distribution that we can evaluate likelihood on. With the evaluated likelihoods, we can arithmetically encode the quantised latent representation ŷ into a bitstream, and assuming that the identical likelihoods are evaluated on the decoding side, the bitstream can be arithmetically decoded into ŷ exactly (i.e. losslessly) (for example, see
We have considered that the latent distribution is most likely a joint distribution with conditionally dependent variables. That is, the variables of ŷ={ŷ1, . . . , ŷN}T have statistical dependencies between each other; they are correlated. As previously visited, with a factorised assumption, the dependency structure is not directly modelled. Hence, if the true latent distribution mŷ(ŷ) does contain statistical dependencies, a factorised assumption on the entropy model pŷ will never attain optimal compression performance (see
By leveraging parametric multivariate distributions, we can capture these statistical dependencies in our entropy modelling if the correlations are modelled adequately. For example, the multivariate normal distribution (MVND), denoted by (μ, Σ), can be used as a prior distribution. The MVND is parametrised by a mean vector μΣ
N and covariance matrix Σ∈
N×N. A comprehensive list of examples of parametric multivariate distributions under consideration for the methods outlined below can be seen in Table 2.3.
However, there are three leading problems with directly incorporating intervariable dependencies in our entropy model:
We have sought to find a remedy to these challenges, and the next subsections will shed light on the methods and technologies that enable or facilitate the employment of parametric multivariate distributions in entropy modelling for AI-based compression. Throughout these subsections, examples are provided of how each method is applied assuming MVND as prior distribution.
In order to take on the challenge of the exploding dimensionality of the latent space, we provide a way to partition the latent space into smaller chunks on which we ascribe intervariable correlations. Ideally, these chunks encompass variables that indeed demonstrate correlative responses, such as locally in the spatial and channel axes (when expressed in array format). By doing so, we prescribe zero correlation for variables that are far apart and clearly have no mutual influence. This drastically reduces the number of parameters required to model the distribution, which is determined by the partition size and therefore the extent of the locality.
It should be noted that the chunks can be arbitrarily partitioned into different sizes, shapes and extents. For instance, assuming array format of the latent space, one may divide the variables into contiguous blocks, either 2D (along the height and width axes) or 3D (including the channel axis). The partitions may even be overlapping; in which case, the correlations ascribed to each pair of variables should ideally be identical or similar irrespective of the partition of which both variables are a member of. However, this is not a necessary constraint.
The effects of the reduced number of parameters required using a partitioning scheme can be understood by an example. Using MVND as an entropy model imposed on latent space ŷ∈N, we can split up the latent space into B contiguous partitions of size m=16 or blocks of 4×4 variables (pixels) along the spatial axes (as seen in the first example in
m and Σ[b]∈
m×m. Whereas a single MVND entropy model on the entire latent space requires
parameters (the second term is because the covariance matrix is symmetric a partitioned latent space with B MVND entropy models require
parameters in total.
Although in this example we have been focused on partitioning of the latent space for tractable dimensionality, the same principle could be applied for any vector space encountered in AI-based data compression.
Depending on the parametric distribution type adopted, the quantity expressing the intervariable dependencies may have different constraints. For instance, the absolute magnitude of the elements in a correlation matrix can never exceed one, and the diagonal elements are exactly one. Some expressions of intervariable dependencies include, but are not limited to, the covariance matrix Σ, the correlation matrix and the precision matrix Λ. Note that these quantities are closely linked, since they describe the same property of the distribution:
Apart from this, all three expressions share common mathematical properties such as symmetry and positive definiteness. Therefore, it makes sense to narrow in on a single expression when discussing the parametrisation of intervariable dependencies. In this case, we will focus on the covariance matrix Σ.
There are multiple ways that we could parametrise Σ whilst satisfying its intrinsic properties. Here are some examples that we have successfully used to date, which are by no means exhaustive.
N+1−i
N×N
Equals length of vector u
Householder matrix
Embedding Householder matrix
Householder reflection of
Example: Suppose our entropy model py over a (partitioned) latent space is an N-dimensional MVND, y˜(μ, Σ). We will assume that Σ is parametrised by its eigendecomposition, the eigenvalues s and eigenbasis B (by Householder). Then, we can perform PCA whitening to decorrelate the zero-centred variables y−μ by transforming with the inverse of the eigenbasis
z=B−1(y−μ)=BT(y−μ)
where z is the decorrelated latent vector. The decorrelated latent variables are now all mutually independent, and distributed as an uncorrelated MVND with eigenvalues as its variances s
z˜(0,I s)
whose probability mass can be evaluated as a joint factorised normal distribution:
To engage with multivariate distributions in an entropy coding setting, we must be able to unambiguously evaluate probability masses. Normally, for simple univariate parametric distributions, there often exists a closed-form expression for the CDF (Equation (2.7)), which provides easy probability evaluation. This is no longer the case for multivariate parametric distributions.
For any continuous probability distribution with a well-defined PDF, but lacking a well-defined or tractable formulation of its CDF, we can use numerical integration through Monte Carlo (MC) or Quasi-Monte Carlo (QMC) based methods. These methods estimate the probability mass over a hyperrectangular integration region Ω⊂N on the N-dimensional PDF py(y; ϕ). These methods rely on uniform sampling of a large number, say M, of pseudo-random or quasi-random perturbation vectors within a zero-centred integration domain, expressed over the dimensions in product form as Ω=Πi=1N[ai, bi]=[a1, b1]×[a2, b2]× . . . ×[aN, bN]. Then, given a sufficiently large sampling size, the probability mass associated with an arbitrary centroid ŷn over the integration domain Ω can be approximated by
where V(Ω)=Πi=1N(bi−ai) is the integration volume over Ω and the perturbation vector is sampled uniformly within the integration boundaries ∈j˜(Ω).
Note that MC- and QMC-based evaluation of probability mass can be done both for univariate and multivariate distributions. This method is also not directly backpropagatable because of the sampling process, however it would be feasible to employ this method in gradient-based training by using gradient overwriting. Furthermore, to avoid non-deterministic probability mass evaluations between encoding and decoding, the same pseudo- or quasi-random process must be agreed upon between either sides of the transmission channel.
In the special case for an MVND, there exists another way of evaluating an approximate probability mass (apart from the PCA whitening approach as explained in previous section) which actually is differentiable. The method will be described in the example below.
Example: A joint distribution p(y) that belongs to the family of MVND, has the property that the conditional distributions of its variables are also normally distributed. That is, the conditional distribution of a variable, given the previous variables, p(yi|y1, . . . , yi−1) is a univariate Gaussian with the conditional parameters ϕi=(
where Ci is the matrix of regression coefficients
and where the inverse operation is the generalised (Moore-Penrose) pseudoinverse.
With the conditional parameters, the probability mass would be estimated in the same way as a univariate normal distribution. Importantly, this formulation is only approximate since the conditioning occurs over a single point, whereas in reality, the probability mass is evaluated over a closed interval on the probability density function. In practice however, as long as the distribution is not converging towards a degenerate case, this method provides a useful approximation for probability mass evaluation whenever Σ is obtained directly and rate evaluation requires differentiability.
We established that multivariate probability density distributions are hard to learn and evaluate with naive methods and require specific approaches to make them work. One of these is using Copula.
In probability theory and statistics, a copula is a multivariate cumulative distribution function for which the marginal probability distribution of each variable is uniform on the interval [0, 1]. Copulas are used to describe the dependence between random variables. In short, Copula is a way to get value from a joint probability distribution using marginal distributions plus a couple-function (→ the Copula). This couple function is there to introduce correlation/dependencies between the marginals.
Let's assume we modelled the latent space with a factorised distribution {Py1, . . . , PyN}, but we want to model it using a joint distribution Py(.). If we assume that our factorised distributions are the marginal distribution of the joint distribution, we can use Copula. Note that this assumption is not always valid, but there are classes of distributions for which this holds, e.g. the multivariate normal distribution, the multivariate Laplace distribution, and others. If we want the joint (multivariate) cumulative distribution function CumPy(y1, y2, . . . , yN)=Prob(Y1≤y1, . . . , YN≤YN), we can express CumPy(·) as a Copula function C with inputs being the evaluation of the marginal cumulative distributions:
CumPy(y1,y2, . . . ,yN)=C(CumPY
Moreover, we can get the density function of the joint distribution by simply differentiating the copula function. Let Py be the joint density function and Py
Py(y1,y2, . . . ,YN)=c(CumPY
The above equation states that the properties of dependence are often efficiently analysed using copulas. An n-dimensional copula is just a fancy name for a joint probability distribution on the unit square [0, 1]2 with uniform marginals.
So what is the Copula function C(·), and how to create it? The copula C contains all information on the dependence structure between the components of (Y1, Y2, . . . , YN). The Copula C(·) is the joint distribution of the cumulative transformed marginals.
The transformed marginals into [0, 1] (probability space):
(U1, . . . ,UN)=(CumPY
The Copula function:
C(u1, . . . ,UN)=Prob(U1≤u1, . . . ,UN≤UN) (2.17)
Let's go through an example to build an intuition behind what Copula is. It is often used to generate correlated random variables of “difficult” distributions. Let's assume we want correlated random variables from a joint (multivariate) hyperbolic distribution. Well, no library can quickly generate these, so what can we do?
If we know the marginal (factorised) distributions of the joint distribution and the correlation that we want, the task is possible. We simulate random correlated variables given the joint multivariate distribution. We transform them to correlated variables in [0,1] using the joint normal distribution's marginals' cumulative distributions. We re-transform these values in a joint hyperbolic distribution by using the inverse marginal cumulative distributions of the joint hyperbolic.
This process is only possible by using the Copula approach.
Our innovation is to use Copula for latent distribution modelling in two ways:
In our world, for everything, there is a dual-representation or duality. For instance, in Physics, we have the Wave-particle duality; for images, we have their representation in the spatial or the frequency domain, and for probability distributions, we have their characteristic functions. Usually, in one world, we can evaluate points easily but struggle with their impact on the surrounding area (particles, spatial domain, probability functions); whereas in the other, we can evaluate their waveform easily but struggle with their impact at a specific point (waves, frequency domain, characteristic functions).
If a random variable admits a density function, then the characteristic function is its dual, in the sense that each of them is a Fourier transform of the other. Let φX(t) be the characteristic function at (wave) position t for random variable X. If the random variable X has the probability density function ƒx(x) and the cumulative distribution function Fx(x), then the characteristic function is defined as follows:
φX(t)=E[eitX]=eitxdFx(x)=
eitxƒx(x)dx (2.18)
The following table summarises the above paragraph. Note that point evaluation in the spatial domain is equivalent to wave evaluation in the wave domain. The wave evaluation in the spatial domain is equal to point evaluation in the wave domain.
Our innovation is to combine our latent density modelling with the characteristic function in multiple ways:
Most of the parametric distributions that have been explored here thus far exhibit unimodality, i.e. their PDF formulation has at most a single peak or cluster. There is nothing that restricts the true latent distribution from being multimodal, or having multiple distinct peaks. In fact, this is especially true for multidimensional latent spaces since signals tend to aggregate into clusters if they carry similar information, and separate from others if the information is dissimilar. This creates a natural proclivity for multimodality of the latent space. If the latent space truly is multimodal, a unimodal entropy model will not be able to model it perfectly.
In order to incorporate multimodality to entropy modelling, it is possible to employ mixture models as prior distribution. A mixture model is comprised by K number of mixture components, which are base distributions either from the same family of distributions or different, including non-parametric families of distribution (see Section 2.4.4). The PDF py is then a weighted sum of each mixture component, indexed by [k]
where π[k]∈[0, 1] represents the mixture weight for the PDF of the kth component py[k]. All mixture components must be defined over the same vector space, and all mixture weights have to sum up to one to ensure a proper probability distribution Σk=1Kπ[k]=1 (which can be done with a simple softmax operation). This implies that a mixture model actually generalises all distributions (see
A main drawback with parametric probability distributions is that they, ironically, impose a prior on the distribution it tries to model. If the distribution type is not compatible with the optimal latent space configuration, the prior effectively stifles the learning process.
We have investigated the generation of non-parametric probability distributions for entropy modelling. Non-parametric probability models are not defined a priori by a parametric family of distributions, but are instead inferred from the data itself. This gives the network many more degrees of freedom to learn the specific distribution that it needs to model the data accurately. The more samples per unit interval, the more flexible the distribution. Important examples are histogram models and kernel density estimation.
There are multiple ways of modelling the distribution without a parametric form. One simple way is to train a neural network l=ƒψ(t), parametrised by network weights, biases and activations p which takes the range of values with non-zero probability as input t and outputs logits l for discrete probability masses for the range. For example, if the quantised latents ŷ∈consist of rounded integers with minimum and maximum (ŷmin, ŷmax) respectively, then t={ŷmin, ymin+1, . . . , ŷmax−1, ŷmax}. The outputted logits would be of the same size as its input ƒψ:
Q|t|→
|t|, where |t| denotes the number of elements in the vector t. To ensure that we obtain a valid probability distribution, the logits must sum up to one either by normalisation
or using the softmax operation
This strategy of learning a discrete PMF can be extended to learning a continuous PDF by interpolating the values between adjacent discrete points (P(yi), P(yi+1)) that are obtained. Extra care must be taken to ensure that the probability density integrates up to one. If linear (spline) interpolation is used, we obtain a piece-wise linear density function whose integral can be easily evaluated using the trapezoidal rule (see
However, this strategy comes with two problems; the range of values t must be finite and known in advance, and the array indexing operations (which is inherently discrete) that is required to infer probabilities does not lend itself well to automatic differentiation frameworks. Hence, another method of interest is learning a 1-D (factorisable) cumulative density function (CDF) which can then be used with Equation (2.7) for rate evaluation. This method relies on designing, parametrising and training a neural network that maps a value directly to a continuous CDF, ƒψ: →[0, 1] which satisfies two constraints:
The first constraint can be satisfied by performing a sigmoid operation
on the return value, or any other range-constraining operation (such as clipping, projection, etc). For the second constraint, there are many possibilities to satisfy this which depends on the network architecture of ƒψ. For instance, if the network is comprised by a composition of K vector functions (convolutions, activations, etc)
ƒψ=ƒK∘ƒK−1∘ . . . ∘ƒ1 (2.20)
its partial derivative with respect to the input, i.e. the PDF pψ, is defined as a chain of matrix multiplications of the Jacobian matrices (which describes partial derivatives with respect to a vector-valued function) of all function components:
pψ=JƒKJƒK−1. . . Jƒ
Without loss of generality, to satisfy the monotonicity constraint, we must ensure that the Jacobian matrix of each function component with respect to the input is non-negative. Examples how this is ensured is by using monotonic (strictly increasing) activation function such as ReLU, Leaky ReLU (with a positive slope), sigmoid and the hyperbolic tangent, and ensuring that all elements in weight matrices of the linear layers are non-negative. Since the method with which the CDF constraints are satisfied varies with the network architecture of ƒψ, the details of how this is implemented is not the important aspect, but rather the step of satisfying these constraint to admit a proper CDF is important.
In this section, we present the following concepts regarding flexible entropy modelling of latent distributions for AI-based data compression with details outlined in the referenced sections. These are considered under the context of entropy modelling of the latent distributions as well as the wider domain of AI-based data compression.
Real-time performance and fast end-to-end training are two major performance requirements of an AI-based compression pipeline. To these ends, we have incorporated fast iterative solvers into a compression pipeline, accelerating both inference, leading to real-time performance, and accelerating the end-to-end training of the compression pipeline. In particular, iterative solvers are used to speed up probabilistic models, including autoregressive models, and other probabilistic models used in the compression pipeline. Additionally, iterative solvers are used to accelerate the inference speed of neural networks.
AI-based compression algorithms have achieved remarkable results in recent years, sur-passing traditional compression algorithms both as measured in file size and visual quality. However, for AI-based compression algorithms to be truly competitive, they must also run in real-time (typically >30 frames-per-second). To date, the run-time issue has been almost completely ignored by the academic research community, with no published works detailing a viable real-time AI-based compression pipeline.
We have however made significant progress towards achieving a real-time AI-based compression pipeline. Here we outline one of our methods for attaining real-time AI-based compression, namely: accelerating AI-based compression using iterative methods for solving linear and non-linear equations. Iterative methods for equation solving improve several aspects of the AI-based compression pipeline. In particular they speed up the execution of Neural Networks, and significantly reduce the computational burden of using various probabilistic models, including autoregressive models.
Moreover, aside from improving inference speeds (leading to real-time performance), iterative methods can significantly reduce the end-to-end training times of an AI-based compression pipeline, which we will also discuss.
First we review iterative methods for solving systems of equations. Suppose we have a set of N variables x1, x2, . . . , xN. Suppose also we have M functions, ƒ1, . . . , ƒM, each of which takes in the N variables and outputs a scalar (a single number). This defines a system of equations
For brevity we can write this system in vector notation. Define the vector x=(x1, . . . , xN) and the vector-valued function ƒ=(ƒ1, . . . , ƒM). Then the system is simply written
ƒ(x=0 (3.1)
A solution to (3.1) is a particular x that, when evaluated by ƒ, makes (3.1) true. Importantly, not all x are solutions. Finding solutions to (3.1) is in fact difficult, in general. Only in very special cases, when the system of equations has special structural properties (such as triangular systems), can a solution to (3.1) be solved exactly, and even then, exact solutions may take a very long time to compute.
This is where iterative methods for equation solving arise. An iterative method is able to compute (possibly approximate) solutions to (3.1) quickly by performing a sequence of computations. The method begins with a (possibly random) guess as to what the solution of (3.1) is. Then, each computation (iteration) in the sequence of computations updates the approximate solution, bringing the iterations closer and closer to satisfying (3.1).
Take for example, the method of fixed-point iteration (sometimes called Jacobi iteration). This method works as follows. An first guess at a solution x0 is initialized (e.g. by drawing
variables randomly, or by setting the initial iterate to zero). Then, for all following iterations t=1, 2, . . . the iterate is set as xt+1=ƒ(x′). Under suitable conditions, the sequence of iterates will converge to a solution of (3.1). The iterations are terminated when the approximate solution is close enough to the true solution (usually measured via the residual ∥ƒ(x′)∥). Fixed point iteration is guaranteed to converge if the function ƒ is contractive (its global Lipschitz constant is less than one).
Fixed point iteration is a very basic method and more sophisticated approaches are available. These include
Each of these methods can be shown to converge to a solution, given particular constraints on the function ƒ. Often these constraints involve calculating bounds on the minimum and maximum eigenvalues of the Jacobian of ƒ. Convexity requirements may be required for convergence to a unique solution, but are not in general needed for convergence to a solution.
In an AI-based compression pipeline, we seek to compress an image or video x=(x1, x2, . . . , xN), where x is a vectorized representation of the image or video. Each component xi of the vector is a pixel of the image (or frame, if discussing videos). To encode the image or video as an image, we need a joint probability model p(x) which measures the likelihood of the image occurring. The filesize of the encoded image is bounded above by the (cross-)entropy of the probability model—the closer the probability model is to the true distribution of images, the better the compression rate (filesize).
However, working with the joint distribution is difficult. Instead, we may exploit the chain rule of probability: the joint distribution is equal to a product of conditional distributions. That is, we will factorize the joint distribution as follows:
Each of the p(xi|x1:i−1) are conditional probabilities. They measure the probability that pixel xi occurs, given that the value of the preceding pixels x1:i−1.
This factorized distribution, as a product of conditional distributions, is in general much easier to work with. This is especially true in image and video compression. When an image is compressed, and sent as a bitstream, it is not the value of the pixels x that is sent, but rather a vector of conditional probability values that is actually converted to a bitstream. This conditional probability vector is defined as
To emphasize, the vector P is the quantity that is actually compressed (by sending it to for example an arithmetic encoder). At decode time, when the image is to be recovered, we must recover x from the conditional probability vector. In other words, we must solve for x from the system of equations
This is an inverse problem, a system of equations that can be solved using one of the iterative methods described above. To make the link to Equation (3.1) clear, we could define the vector valued function ƒ as the vector of conditional probability functions minus {circumflex over (p)}. Then the system of equations is in the form of (3.1):
Note that system (3.4) has a triangular structure: the i-th conditional probability depends only on the value of the previous variables. This makes it particularly easy to solve, especially using the Jacobi iterative method (fixed point iteration). In fact, with an autoregressive model, the Jacobi iterative method is guaranteed to converge to the true solution in at most N steps. In practice however, an acceptable approximate solution can be achieved in significantly fewer steps, depending on the tolerance threshold E (refer to Algorithm 3.1).
Triangular systems can also be solved serially, one equation at a time. In a linear system, this is called forward substitution (backward substitution). In a serial solution method, first x1 is solved from the equation p (x1)={circumflex over (p)}1. Then, x1 is substituted into the equation p(x2|x1)={circumflex over (p)}2, which is then solved for x2. Both x1 and x2 are substituted into the third equation, which is then solved for x3. The process is continued serially through all equations until finally the entire vector x is recovered.
Unfortunately, this serial process is very slow. It requires exactly N steps, and cannot be done with any fewer calculations. Contrast this with an iterative method, which can converge to an acceptable solution in significantly fewer than N iterations. Moreover, the serial procedure's computations are applied one element (pixel) at a time. In contrast, the iterations of the fixed point scheme (or any iterative method) are applied to the entire image, and can exploit parallelization routines of modern hardware (such as Graphics Processing Unit or a Neural Processing Unit).
What form do the conditional probability functions p(xi|x1:i−1) take? We now review types of autoregressive probabilistic models that may be used in a compression pipeline. One useful approach is to model the probability function is with a basic 1-dimensional probability function with parameters θ. The parameters θ will be the functions of the preceding x1:i−1 variables. So for example, we could model p(xi|x1:i−1) with the Normal distribution
p(xi|x1:i−1)=N(xi;μ(x1:−1),σ(x1:i−1)) (3.5)
Here the mean parameter μ and the variance parameter σ are the output of functions of x1:i−1. In an AI-based compression pipeline, typically neural networks are used for these functions.
There are many possible choices of autoregressive models that can be used to encode the variable into a bitstream. They are all variants of the choice of function used to model the conditional probabilities. The following is a non-exhaustive list. (In the following examples we use the Normal distribution as the “base” distribution, but any distribution could be used)
where each Lk and Mk are filter-bank models (possible linear functions).
Note that the discussion up until this point has been focused on using autoregressive models for probabilistic modelling on an input image x. However, there are many other variables that autoregressive models can be used on:
Although conditional probability distributions are a main component of the compression pipeline, Deep Render still has use for joint probability estimation (estimating the unfactorized joint probability p(x)). This can be done using a Normalizing Flow (refer to our PCT patent “Invertible Neural Networks for Image and Video Compression”, and for a discussion of use-cases). Recall that a joint probability distribution can be estimated by a change of variables ƒ: x∈N
z∈
N:
Here (ƒ(x); 0, I) is the standard multivariate normal distribution, and
is the determinant of the Jacobian of the transformation ƒ.
Typically, ƒ is constructed to be easily invertible, and also to have a tractable determinant formula. This can be done using an autoregressive model. The function ƒ could be made of a series of transformations: ƒ(x)=ƒN∘ƒN−1∘ . . . ∘ƒ2∘ƒ1(x). Each of the ƒi's has an autoregressive structure:
ƒi(y)=g(y1:i−1;θi) (3.11)
where y is the input to the function at the i-th composition. In other words, the function at the i-th place in the chain of compositions only depends on the preceding i−1 variables. The function g could be any function parameterized by θ that is invertible (bijective). So described, this is an Autoregressive Flow.
The process of inverting an autoregressive flow is to solve the system
ƒ(x)=z (3.12)
for x given z. This can of course be done with an iterative solver, and in particular, since the system is triangular (autoregressive), it can be solved easily with fixed-point iteration (Jacobi iteration). Note that in an autoregressive flow, the computing the forward map z=ƒ(x) is typically quick and computationally easy, whereas inverting the system (3.12) is hard and computationally more difficult.
An example where the opposite is true is the inverse autoregressive flow. In this setup, the inverse function ƒ−1(z)=x is modeled as a composition of functions. ƒ−1(z)=ƒ1−1∘ƒ2−1∘ . . . ∘ƒN−1−1∘ƒN−1(z). Each of the inverse ƒi−1's has an autoregressive structure:
ƒi−1(y)=g(yi:i−1;θi) (3.13)
Again, the function g should be bijective so that it can be inverted. In this case the change of variables formula is
And now generating x from z is easy, whereas finding z from x is difficult and involves solving the system
ƒ−1(z)=X (3.15)
This system can be solved using an iterative solver.
One possible variant of the normalizing flow framework is to define the composition of functions as infinitesimal steps of an continuous flow. In this setting the final variable z is the solution to an Ordinary Differential Equation i=ƒ(z; θ) with initial condition z(0)=x. The function ƒ may have an autoregressive structure. Continuous normalizing flows are appealing in that they are easily inverted (by simply running the ODE backward in time) and have a tractable Jacobian determinant formula.
The bulk of this section has focused on autoregressive models, their use in compression pipeline, and how they define systems of equations that can be solved using iterative methods. However, many of the autoregressive methods can be generalized to non-autoregressive methods. This section will illustrate some non-autoregressive modeling tasks that can be solved using iterative methods.
Rather than modeling the joint distribution p(x) as an autoregressive factorization of (autoregressive) conditional probabilities, we may simply model the conditional probabilities explicitly from a defined joint distribution.
For example, suppose we model the joint distribution with a standard multivariate distribution, such as the Multivariate Normal Distribution.
Here Σ is the covariance matrix and p is a mean vector. The constant Z is a normalizing constant so that the RHS has unit mass.
The conditional probabilities are defined via the following formula (here xi=(x1, x2, . . . , xi−1, xi+1, . . . , xN) is the vector missing the i-th component)
The denominator is obtained by marginalizing out the i-th variable. Notice that the conditional probability model here depends both on past and future elements (pixels). This is a significantly more powerful framework than an autoregressive model. Notice also that integration constants cancel here. So for example, with a Multivariate Normal Distribution, the conditional probability density is
The denominator here has a closed form, analytic expression, and so the conditional probability is simple to evaluate. In a compression pipeline, under this framework, to encode a variable x we would construct a vector of conditional probabilities {circumflex over (p)}, using the tractable formula for conditional probabilities (either (3.16) in general, or (3.17) if using Multivariate Normal). Then, at decode time, the vector x is recovered by solving the system
for x using an iterative method.
The parameters of the joint distribution (such as for example the precision matrix Σ−1 and the mean μ) can be produced by a function of side (or meta-information) also included in the bitstream. For example we could model the joint distribution as
where z is the side-information variable, and μ(·) and Σ−1=M(·) are the outputs of functions (possibly neural networks) of the side-information.
Rather than modeling the joint probability distribution with a “standard” multivariate distribution, we can model the joint distribution with a Markov Random Field. A Markov Random Field (sometimes called a Gibbs distribution) defines a joint probability distribution over a set of variables embedded in an undirected graph . This graphical structure encodes conditional dependencies between random variables. So for instance, in an image, the graph variables could be all pixels in the image, and the graph vertices could be all pairwise adjacent pixels.
Contrast this with autoregressive models: autoregressive models are defined on directed acyclic graphs; wherease Markov Random Fields are defined on undirected (possibly cyclic) graphs. Essentially, a Markov Random Field is a rigorous mathematical tool for defining a joint probability model that uses both past and future information (which is not possible with an autoregressive model).
The unnormalized probability density (sometimes called a score) of a Markov Random Field can be defined as
Here cl() are the cliques of the graph. In a graph defined on an image, with edges between pairwise pixels, the cliques are simply the set of all pairwise adjacent pixels. The definition of a clique is well know in the field of graph theory, and is defined a subset of vertices of a graph such that all variables (vertices) of the clique are adjacent to each other. The functions ϕc are called clique potentials. Often they are defined via an exponential ϕc(xc)=exp(ƒc(xc)). In our compression pipeline, the functions {ƒc} could be for example quadratic functions, neural networks, or a sum of absolute values. The functions ƒc could be parameterized by a set of parameters θ (which may be learned), or the parameters could be the function of some side information.
The joint probability density function is defined by normalizing (3.19) so that it has unit probability mass. This is typically quite difficult, but since in compression we are mainly dealing with conditional probabilities, it turns out this normalization constant is not needed.
To illustrate how conditional probabilities are calculated, let's consider a simple graph of four random variables (A, B, C, D), with edges {(A, B), (B, C), (C, D), (D, A)}. Note that in this example the cliques are just the edges. The score function is {tilde over (p)}(a, b, c, d)=ϕ1(a, b)ϕ2(b, c)ϕ3(c, d)ϕ4(d, a). The conditional probability, say p(a|b, c, d) is given by
since cliques without vertex A cancel, and the integration constant cancels as well. Thus conditional probabilities can be easily calculated with an analytic expression, provided the integral in the denominator is tractable.
Therefore, just like with an autoregressive model, Markov Random Fields can be used to encode a variable x via a conditional probability vectors. And, just like with an autoregressive model, the variable x may be reconstructed at decode time by solving a system of equations for x in terms of {circumflex over (p)}. Just like an autoregressive model, the variable to be encoded need not be an image, but could be a latent variable, or could model temporal frames in a video (or latent variables of a video).
We remark that other probabilistic quantities can be easily derived from Markov Random Fields using iterative methods. For example, the marginal probabilities can be obtained using belief propagation, and other message passing algorithms, which are specific iterative methods designed for Markov Random Fields.
The conditional probabilities need not be modeled explicitly from a known joint distribution. Instead, we may simply model each of the conditional probabilities via a function ƒi: N
[0, 1]. The vector valued function is defined as ƒ=(ƒ1, . . . , ƒN). Each of the functions ƒi could be parameterized via a parameter θ, such as in a neural network. Then on encode the conditional probability vector is calculated as {circumflex over (p)}=ƒ(x; θ). The function ƒ may depend on side information z also encoded in the bitstream. Then, on decode, the variable x is recovered by solving the system {circumflex over (p)}=ƒ(x; θ) for x. This approach is sometimes called a Dependency Network.
This process could be inverted, so that a system is solved iteratively at encode time. Then at decode time, the variable {circumflex over (x)} may be recovered quickly without using an iterative solver. In this setup, we define a bijective function g: [0, 1]NN. At encode time, the conditional probabilities are given by solving the system g({circumflex over (p)})=x for {circumflex over (p)} using an iterative solver, given an image x. (Essentially inverting {circumflex over (p)}=g−1(x). Then, at decode time, the variable is reconstructed by simply calling the function x=g({circumflex over (p)}).
Finally, we note that iterative solvers need not be used only for probabilistic modelling. In fact, iterative solvers can be used to decrease execution time of neural networks themselves. The execution path of a feed-forward neural network itself has a triangular (autoregressive structure). For example, let x0 be the input to the first layer of a neural network. Let ƒ1, . . . , ƒL be the layers of a neural network. Then the output y of a feed forward neural network is given by the following non-linear autoregressive (triangular) system
Notice that this system is triangular (autoregressive): each line depends only on the preceding variables. Therefore, a solution can be given by using an iterative method tailored to autoregressive structures, such as fixed-point (Jacobi) iteration. In practice, we have found that this approach can lead to significant speed ups in inference and training times.
Differentiation and training may be accomplished using any of the methods discussed in the next section.
Using iterative methods inside an end-to-end compression pipeline has numerous advantages.
Among the foremost advantages is a reduction in training times. For example:
However using iterative methods inside a neural network presents some challenges, especially in regards to end-to-end training of a compression pipeline. We have taken a number of steps to alleviate these problems. The main challenges (and their solutions) of end-to-end training with iterative solvers are the following.
In end-to-end training of a compression pipeline with an iterative solver, we must compute gradients of the solutions outputted by the iterative solver. There are several ways to do this:
Then, from basic rules of calculus, we first use implicit differentiation on the system:
The unknown variable in this system is
It can be solved for using an iterative solver (while the expression
is a Jacobian-vector product and can be easily evaluated with automatic differentiation). Once a solution is found, then it is dropped in, via the chain rule, to calculate
Often in an AI-based compression pipeline, the variable to be solved for in a system of equations is a quantization of another variable. However, during training, it is not feasible to access (calculate) the quantized variable—it would simply take too long, making training unfeasible. Typically the quantized variable in question is a quantized latent ŷ=Q(y). This is the problem of accessing the ground-truth quantized latent during training. Several approaches have been developed to overcome this problem during our training, including:
All iterative solvers in this document can be adapted to solve for quantized variables, if during training the solvers given access to a simulated (approximate) quantized variable. Of course, ideally the ground-truth quantized latent would be used, but in general this is difficult, and remains an active area of research.
In AI-based compression, the rate and distortion are the two main objectives we aim to optimise. The rate aims to make the message we are streaming as small as possible in size (bits), while the distortion aims to keep the fidelity of the received message as close to that of the sent message. Translating this to the transmission of an image, the sender encoding the image using the codec, hoping to reduce it's file size as much as possible, streams it to the receiver, who decodes the image and hopes that the quality of the image is as good as the original. However, these two aims of reducing the file size and maintaining the quality are at odds with each other. Reducing the file size of an image makes the quality of the image worse (lossy compression).
There are multiple ways to define a distortion/fidelity metric in AI-based training, and the only requirement they have is that they be smooth and differentiable in order for us to be able to differentiate them with respect to their inputs. This makes training our AI-based compression pipeline feasible. Along with this, another aspect that has recently been considered important for a distortion metric is that it must be tuned to the human visual system. In other words, differentiability is not the only criteria for our distortion metric, but it must now take into account the human visual system. Asking a mathematical function to take into account the human visual system is impossible currently, as it firstly assumes we understand how humans perceive images (what they prefer in an image and what they discard), and secondly that we can build such a complex function in a differentiable way.
The method aims to solve this problem by learning a function that takes as input a distorted and ground truth (GT) image, and outputs a score which indicated how a human viewer would perceive the image (1 is poor quality, 10 is indistinguishable from GT). A requirement is that we have some human labelled data to teach our function. Furthermore, we outline some training strategies and methods to enhance our results.
Ultimately, the function learnt, called Deep Visual Loss (DVL) acts as the distortion metric and is used to train a compression pipeline for image and video compression.
We learn to approximate the human visual system in a supervised fashion, where we define a function ƒ and subsequently teach it to fit the human labelled data. For this learning process, we must first acquire the data. In this section, we outline some methods to acquire the data.
The primary method for acquiring data is through human labelling. Here, we collect a wide variety of images across different quality levels and present them to humans and ask them to assess the quality using one of the following methods (these methods are well understood and commonly used in literature of human quality assessment):
In these test, we ask candidates to select the preferred image or rate an image on a scale of 0 to 5, which gives us a label per image. We do this over thousands of candidates and images (to get statistical significance) and use statistical methods such as Z-score and extreme value analysis to reject outliers. The result of this is a collection of human labelled images.
A key component of the data acquisition process is collecting the distorted image samples humans will assess the quality of. These samples have to be representative of what will be seen when the compression pipeline is being trained. To understand this intuitively, think of the function as a mapping from an image to a value. If the input image has previously been seen during the training of this function, we are able to perform the mapping from image to value accurately. However, if the image it too dissimilar from what was used to train our function, the mapping can suffer from inaccuracies, ultimately leading to difficulties in the training of our compression pipeline.
To mitigate this, we ensure our dataset used to train our function includes a wide range of distortions and mainly, distortions introduced using AI-based compression encoder-decoder pipelines. This is done through simply forward passing a set of images through a trained AI-based compression pipeline. Alternative, it is also possible to saves images at different time steps of an AI-based compression pipeline training, as this will provide better coverage of images we are likely to see. When saving images during the training of a pipeline, we propose to use all existing distortion functions.
From herein, this data consisting of images of different qualities and their respective human labels will be referred to as acquired data or human labelled data (HLD).
In this section, we will detail the methods used to learn ƒ from HLD. Since we have image data and a value to map to, there are many methods that can be used here. We outline the details of neural networks and regression based methods.
We propose to use neural networks to learn from the HLD. We refer to this network as a Deep Visual Loss (DVL) network. Neural networks are termed as universal function approximators, which essentially means that given a neural network with enough parameters, we can model an arbitrarily complex function.
This makes them attractive as function approximators. There are many configurations we can use when it comes to defining this neural network, and our claim does not limit us to any particular configuration. However,
In
Once we have defined such a network, we train it using the HLD in a supervised training scheme using standard and widely known deep learning methods such as (but not limited to) stochastic gradient decent and back propagation.
As mentioned above, we train our Deep Neural network on HLD to predict the labels of HLD, which gives us an indication of how a human would rate them image. In this section, we outline some methods to improve our training. The pseudo-code shown in Algorithm 4.1 below shows how the training scheme may look like.
The data acquisition stage is expensive, especially if we want to get a sufficient amount of data and capture a wide range of distortions. It is also the case that the more data deep neural networks have for training, the better they perform. We provide an automated method to generate labelled data, which is used to pre-train our DVL network before it is trained on HLD. It is widely acknowledged that pre-training can help with learning and generalisation. In order to generate this data for pretraining, we use bit allocation (rate) as a proxy for perceptual quality. During this method, we generate the labels for our distorted data using the bit-rate. Our AI based compression pipeline can be conditioned on or trained for several lambda values. These values determine the trade-off between the rate (bits allocated to the image) and distortion (visual quality). We use a range of lambda, from low to high, to generate distorted images. For our case, higher lambda values generate visually pleasing images while lower lambda values generate visually distorted images. We can pair these lambda values with an appropriate visual quality value, giving the lowest lambda 1 and the highest lambda value of 9, and the ground truth value of 10. Here, 10 represents the best visual image and 1 represents the worst image.
This method provides us with a plethora of labelled data, without the need for human evaluators. This labelled data can be used to train and pre-train our DVL network.
We propose to make DVL multi-resolution.
where N is the number of resolutions.
We enhance the training of DVL through initialising and training multiple networks on the data separately. This method is generally referred to as an ensemble of networks in literature, and it makes the predictions more robust, since each of the DVL is randomly initialised, and will find a different minimum on the loss surface. Therefore, averaging the results of these various instantiations, has the affect of increasing robustness through decreasing variance and ignoring outliers.
Apart from random initialization of the same network, we use multiple models with varying architectures in our ensemble. This is known as model variation ensembles.
During the training of an ensemble of these network, we compute the loss of each network separately using its output score s and the respective GT value h. However, during inference, we use the average result.
Depending on the data acquisition method, we acquire different formats of training data labels. For examples, when considering single and double stimulus tests, we will receive a score for each image, between 0-5 (where 0 represents bad quality, and 5 represents good quality). When considering alternative forced choice, we will get a binary output, showing which image is superior.
Training of DVL network can be performed on any one of the data acquisition methods. To learn on 2FAC data, we are able to convert the 2FAC rankings into per image score (using methods existing in literature such as Thurstone-Mosteller or bradley terry), which the DVL network can regress. Alternatively, we can also employ a method by which we feed in all three images of the 2FAC into a network, asking the network to predict distances for each, which we send into a fully connected network to predict the result of the 2FAC.
Here the blue and green convolution blocks share weights, and once the network is trained, we can use the score s to train our compression pipeline.
Besides using neural networks directly on the images to predict visual loss scores, we are also able to use a weighted mixture of multiple existing loss functions to predict the visual loss score. When employing these methods, we refer to the visual loss score as DMOS.
Specifically speaking, we provide an aggregate visual loss function which is based on a set of individual distortion loss metrics, each of which is evaluated for the distorted image with reference to the original image and multiplied with a coefficient before being summed together. The coefficients are found by regression analysis between the individual distortion losses and subjective opinion scores, ensuring that the final visual loss score correlates highly with HLD. The following sections will act as a high-level description of the regression based visual loss function.
Given a GT image x, its distorted counterpart z, an enumerated set of N different distortion loss functions {Li}i=1N(outlined in a later section) and a set of regressed (polynomial) loss coefficients {{pij}j=0m}i=1N and an intercept value C, the DMOS loss can be expressed as a sum of polynomials
The individual loss functions {Li} utilised in the DMOS include, but are not limited to, the following:
The aforementioned losses are so-called full-reference image quality assessment algorithms, which means the distorted image is compared to its reference counterpart. However, DMOS is also intended to incorporate no-reference image quality assessment algorithms, including, but not limited to:
The coefficients {Ci}i=0N are optimised for using various types of regression analysis against HLD. The goodness-of-fit is assessed by computing various correlation coefficients, such as Pearson, Spearman and Kendall rank correlations, as well as root mean squared error (RMSE). The types of regressions may be used singularly or in combination with each other and include:
One of the provided methods above is to apply Bayesian methods (Bayesian linear regression & Gaussian process regression) in a similar fashion as described above. The key here is that we get an uncertainty measure with each prediction. This uncertainty measure indicated how certain the model is about a particular prediction. This allows us to modify how we update our compression network. For example, If we are really certain that our prediction of the visual loss score is correct, we use the gradients to update our compression network, however, if we are not sure, we can skip that gradient step since it is likely to be incorrect information. This is particularly useful when there is not a lot of data available as it is more likely that the model will encounter samples it is uncertain about.
Finally, once we have a function ƒ trained, our objective is now to train our compression pipeline through using this ƒ. Below we outline how we do this and some additional training strategies.
When training our compression pipeline using ƒ, we add additional terms to the distortion loss such as but not limited to MSE. These additional terms, which are computed as using the GT and predicted images, are used along with the visual loss score obtained from one of the methods above to train our compression pipeline. This has the benefit of acting as a regulariser to the visual loss learnt from HLD. For example, in the case where visual loss is uncertain, this regulariser loss steps in to provide gradients for the compression network that are still meaningful. We propose to use any combination and number of losses here, for examples, one possible combination is to DMOS using deep pre-trained features whose weights are learnt using linear regression along with PSNR. Another alternative is the use of DVL network with MSE or PSNR.
Adding this additional loss term also helps with stability in training of the compression pipeline using ƒ.
Rather then training using the learnt ƒ from scratch, we train our network using MSE or another distortion function initially for some number of iterations, and introduce ƒ in slowly when the network has stabilised. This method help stabilise training.
It is possible for us to skip the gradients of images we are uncertain about, or give them a low weighting. This can be done through using the uncertainty measure output by Bayesian methods, where is the uncertainty value σ2 is high, we can skip the gradients from the distortion, or specifically, ƒ.
We use this method in AI based image and video compression as well as niche markets such an AR/VR, self driving cars, satellite and medical image and video compression.
We learn a function from compression specific human labelled data to be used as part of the distortion function for training AI based compression pipeline.
Quantisation plays an integral role in compression tasks and enables efficient coding of the latent space. However, it also induces irreversible information losses, and encumbers gradient-based network optimisation due to its uninformative gradients. The causes for the coding inadequacies due to quantisation and the methods with which we can alleviate these, including innovations and technologies, are described.
Formally, for input data of any type (image, video, audio, text, etc.), data compression is the task of jointly minimising the description length of a compact representation of that data and the distortion of a recovered version of that data. In effect, we have two terms, the rate R and the distortion D, that we simultaneously are finding a minimum for as means of a weighted loss sum and a trade-off parameter A that describes the relative weighting of each term:
=R+λD (5.1)
Whether it is the traditional or modern kind, the idea of lossy compression is nevertheless incredibly likely to host some form of discretisation procedure. This is called quantisation, which entails the mapping of a value from a large set, say, the number 3.1415926536 (if the set is all multiples of 0.0000000001), and assigning it to one of many pre-determined states, say the number 3.14, from a countably smaller set (multiples of 0.01). Naturally, there are many, many more states of the former set than the latter (more exactly, 100 million times more).
Quantisation has strong implications on the compression objective as a whole, especially in the latent space where it is applied. With fewer states to consider, the ability to describe a state from the quantised set is more convenient from an information theoretical perspective. This facilitates the task of reducing the description length of the compressed bitstream, or simply put the rate term. On the other hand, assuming that the original state contains very particular information about the source data, quantisation irrevocably discards some of that information as a consequence. If this information cannot be retrieved from elsewhere, we cannot reconstruct the source data without inducing distortions in our approximation.
Perhaps the most typical form of quantisation in this context is rounding. In compression, it manifests in the rounding of latent variables in the bottleneck, for instance to the nearest (even) integer point, such that the latent space is discretised into finite states which lend itself well to entropy coding. Since entropy coding is an integral component of the compression pipeline, we cannot get away without incorporating quantisation under this framework.
The problem, in particular for AI-based compression, is that gradient-based network training requires differentiability through all the relevant operations. Unfortunately, this gives rise to incompatibilities with quantisation operations, since most do not have useful gradients that are necessary for backpropagation. For example, as seen in
The focus here is to
See section 2.2 for a detailed section on mathematical preliminaries.
In this section, we justify the vital role of quantisation in lossy image compression from the viewpoint of the latent space. We characterise the different types of quantisation, and talk about how quantisation impacts a gradient-based network training process. We present some quantisation strategies known to existing literature, and draw parallels with the variational inference framework of posterior distribution matching. Lastly, we introduce the three gaps of quantisation and their implications in neural networks for AI-based data compression.
The latent vector y∈M, or just latents, acts as the transform coefficient which carries the source signal of the input data x. It is often, but not necessarily, retrieved from an analysis transform of the data, y=ƒenc(x); hence, the information in the data transmission emanates from the latent space.
The latents generally consists of continuous floating point values. However, the transmission of floating point values directly is costly, since the idea of entropy coding does not lend itself well to continuous data. Hence, it is possible to discretise the latent space in a process called quantisation Q: M→
QM(where
QM denotes the quantised M-dimensional vector space,
QM⊂
M). During quantisation, latents are clustered into predefined bins according to their value, and mapped to a fixed centroid of that bin (such as rounding to nearest integer). We normally denote quantised quantities with a hat symbol, such as ŷ.
From a probability perspective, the consequence of a discretisation process is that a continuous density collapses to a discrete probability mass model. For instance, if the latent variable yi is distributed according to a continuous probability model py
The effect of quantisation on the assigned quantisation task is dual. On one hand, the set of possible values for the latents is reduced significantly, allowing for compatibility with entropy modelling which enables shorter descriptors of the latents. On the other hand, the process induces irrecoverable information loss, which significantly limits expressiveness of the latent variables. This manifests in corruptions and artefacts in the reconstructed image obtained from the decoder {circumflex over (x)}=ƒdec(ŷ). Very broadly speaking, the coarseness of quantisation (for instance, the width of the bins) has the capacity to determine the rate-distortion tradeoff levels. The coarser the quantisation, the lower the bitrates achievable but the larger the distortion. The effects are reversed for finer quantisation.
In this section, we clarify the distinction between scalar and vector quantisation given an M-dimensional vector with continuous elements y=[y1 y2 . . . yM]T. Scalar quantisation trivially means that each element is quantised individually as if they were scalars, without regard for other elements in y. Each element can have Ci arbitrary centroids, but the centroids pertain to their own dimension only, ŷi∈{ŷi[c]}c=1Ci⊂, ∀i∈[1 . . . M]. Thus, quantisation entails a mapping to the nearest centroid in each dimension.
Vector quantisation, on the other hand, is a mapping to centroids with explicit multiple dimensionality. It considers partitions (or the entirety) of the vector, {y[b]}b=1B where B is the number of vector partitions. Each partitioned vector is quantised to a set of Cb centroids, ŷ[b]∈{ŷ[b],[c]}c=1Cb⊂|y
Moving forwards, when discussing the methods and technologies here, the scalar quantisation framework will be assumed. Nevertheless, all of the presented methods and technologies work equally well with a vector quantisation assumption, and all the concepts pertaining to these should also encompass their extension to vector quantisation.
Conventionally, the optimisation or training of a neural network g(·) is done according to the principle of the chain rule. It assumes a composition of differentiable functions ƒk during forward propagation of the data to evaluate the output, conventionally a loss metric
g=ƒK∘ƒK−1∘ . . . ∘ƒ1 (5.2)
where each function outputs a hidden state hk which acts as the input for the next function:
hk=ƒk(hk−1) (5.3)
Gradient-based optimisation of neural networks relies on computing the first-order gradients of some loss function =g(x). These gradients then flow backwards through the differentiable operations in a process called backpropagation by virtue of the chain rule (making the independent variable a scalar for visibility)
where
is simply the derivative of ƒk with respect to the input. The gradient signal cascades backwards and updates the learnable network parameters as it goes. For this to work effectively, the derivative of each function component in the neural network must be well-defined. Unfortunately, most practical quantisation functions have extremely ill-defined derivatives (see
We start by exploring whether we would be able to replace gradient-based optimisation. Given a set of M continuous latent variables, assume that the optimal quantised latent configuration is retrieved by either rounding up or down to the nearest integer. This task can be formulated as an integer or discrete optimisation problem, which is clearly intractable (a so-called NP-hard problem): the possible evaluation points would be of order 2M, where M is already a large number (in the order of 104 to 105). Consequently, it seems most likely that we have to find workarounds for the non-differentiability property of quantisation functions.
Can we make quantisation differentiable? It certainly seems so; let us take the example of the integer rounding function as quantisation function, and rewrite it like this:
Q(yi)=└yi┐=yi+(└yi┐−yi)=yi+ε(yi) (5.5)
Here, we have defined the function ε: →[−0.5, +0.5] as the quantisation residual, the difference between the quantised and unquantised variable (for integer rounding). Under these circumstances, the quantisation residual is limited in magnitude, and can be seen as an additive term to the original input. Hence, we can model the effects of quantisation with a quantisation proxy
{tilde over (Q)}(yi)=yi+εi (5.6)
where εi is no longer input-dependent but is rather a noise vector sampled from an arbitrary distribution, such as a uniform one, εi˜(−0.5, +0.5). Since we do not need gradients for the sampled noise, we can see that this quantisation proxy has a well-defined gradient:
From here on in, it becomes convenient to distinguish between true quantisation operations and quantisation proxies. The former refers to operations that actually discretises the space, making it convenient for entropy coding and other desirable properties during inference and deployment. The latter refers to differentiable stand-in functions that mimic the behaviour of the discretisation process, whilst retaining a continuous space to allow for network training or applications where gradient propagation is required.
Indeed, automatic differentiation packages allow for customisation of backward functions. In other words, we could define a functional expression for
that would allow gradients to pass through in any desired manner. This is called gradient overriding which also has the ability to form valid quantisation proxies.
Data compression is related to a variational inference framework, through its aim of minimising the Kullback-Leibler (KL) divergence between the true posterior distribution pθ(ŷ|x) and an approximate variational distribution qϕ(ŷ|x) (obtained from the encoder after quantisation). The optimisation problem is traditionally posed as such
where the expectation is taken over the empirical data distribution px(x) and ϕ and θ are the parameters for the inference and generative models, respectively. The objective function in Equation (5.8) can be expanded to form a sum of loss terms:
It can indeed be shown that each of these terms relate to specific loss terms occurring in data compression. For instance, the likelihood term (the second one) log pθ(x|{tilde over (y)}) is related to the distortion or reconstruction loss, and the differential entropy term (the third one) represents the encoding costs of {tilde over (y)}. The last term log pθ(x) is simply the marginal distribution of the observed data, which we cannot influence; hence, we can drop this term from the scope of optimisation.
The focus for quantisation falls on the first term, log pθ({tilde over (y)}|x), which is the logarithm of the conditional distribution of the approximate latents given the data input. Since an encoder normally maps our input to raw latents, y=ƒenc(x), the conditional distribution becomes dependent on the quantisation that is imposed on the latents. In many instances, independent uniform noise quantisation is assumed as the quantisation proxy which yields the following property for qϕ({tilde over (y)}|x):
It can be argued that since a uniform distribution with unit width has constant probability density of 1, the logarithm of this density evaluates to zero. However, this is not a trivial assumption; without this assumption for your quantisation proxy, this term cannot be ignored. Our studies indicate that the noise distribution is neither factorised nor symmetric, and may indeed be highly context-dependent. We mean that the distribution of quantisation residuals contain certain statistical dependencies, which suggests that if we are able to model them during training, the optimisation process would imitate the compression task with true quantisation more closely.
In the field of AI-based data compression where gradient-based optimisation and quantisation both play integral roles, but are mutually incompatible, the discretisation process introduces certain inadequacies that manifest as differences between the ideal case and practical case.
We can identify and characterise three such gaps:
Under many deep learning applications, it is ideal to work with continuous (real) values. Since a discretisation process such as rounding breaks the continuity, it does not lend itself well to gradient-based network optimisation. One way we could remedy this is by substituting the true quantisation operation Q(·) with a quantisation proxy {tilde over (Q)}(·) during network training. In inference and deployment, we do not require differentiability, during which we would revert back to using Q(·). Thus, the discretisation gap refers to the misalignment in the outputs ŷi and {tilde over (y)}i, produced by Q(·) and {tilde over (Q)}(·), respectively. An example of a quantisation proxy that yields a discretisation gap is noise quantisation, {tilde over (Q)}(yi)=yi+εi={tilde over (y)}i where εi are random samples drawn from an arbitrary noise source (the selection of the noise source is a design choice). While it is intended to simulate the effects of true quantisation such as rounding, Q(yi)=└yi┐=ŷi, it is clear that in general {tilde over (y)}i≠ŷi.
Since the loss function consists of two components, the rate R and distortion D, both of which is dependent on the quantised latent variable, the misalignment in the quantisation output propagates onward to each of the loss component. Crucially, the quantised latents conditioning each component do not need to be the same. Past the quantisation function, the algorithm branches out where, on one hand, the entropy model computes the rate term from the first version of the quantised latents ŷ[R], and on the other hand, the decoder (or hyperdecoder) admits the second version of the quantised latents ŷ[D]. This implies that we, in fact, have two discretisation gaps to consider for each set of latents (see
Most entropy coding schemes are only defined for discrete variables, and therefore require a discrete probability model (or entropy model). The problem is that discrete probability models do not provide useful gradients for continuous inputs. Usually, a continuous relaxation of the entropy model is adopted, for instance by employing uniform noise quantisation as quantisation proxy. If the true quantisation is integer rounding, uniform noise quantisation with noise sampled from (−0.5, 0.5) has the property that the resulting continuous density coincides with the discrete probability distribution.
(−0.5, +0.5) gives rise to a continuous relaxation of the PMF Pŷ
(−0.5, +0.5), and coincides with all values of the PMF.
However, the differences in the character of the probability distributions give rise to misalignment in the likelihood evaluation between the continuous and discrete models. This inadequacy is termed the entropy gap.
Although the entropy gap might seem related to the discretisation gap, there are a couple of fundamental differences. Most importantly, the discrepancy manifests itself in the evaluated likelihood for the rate term, where the continuous approximation will in most cases underestimate this quantity. Secondly, whilst the discretisation gap pertains to both the rate term and distortion term, the entropy gap only concerns effects on the rate.
The gradient gap arises when the gradient function of the assumed quantisation proxy has been overridden with a custom backward function. For instance, since the rounding function has zero-gradients almost everywhere, the STE quantisation proxy {tilde over (Q)}(·) assumes its derivative to be equal to be one, such that
(see
For every quantisation proxy that is defined with its own custom backward function that is misaligned with the forward function's analytical derivative, the gradient gap is manifested.
We have been very prolific in pushing the frontiers of quantisation for AI-based data compression by our rigorous development of theory and experimental tests. This section introduces a range of innovations. These are all presented thematically in their own individual subsections below.
Choosing the family of parametric distributions for the entropy model may at first glance appear to be detached from quantisation. However, as shall be seen momentarily, the choice of the parametrisation for the entropy model assumed for the latent distribution py matters a great deal for quantisation, especially with regards to eliminating gradient biases that arise from the quantised variable.
Consider the rate loss function of the continuously relaxed likelihood p(yi; ϕi) which is a cross-entropy term=−log2p({tilde over (y)}i;ϕi) (5.11)
where the likelihood is often evaluated as a difference in CDFs of half a quantisation bin's distance from the evaluation point,
Differentiating R with respect to {tilde over (y)}i (and assigning p{tilde over (y)}
Using the fact that the gradient of the CDF is equal to the PDF,
we obtain
For a univariate Laplacian distribution, the PDF p({tilde over (y)}i; μi, bi) and CDF F({tilde over (y)}i; μi, bi) have the analytical formulae
Assuming integer quantisation (Δi=1.0) and plugging in these equations into Equation (5.14), we get
where {tilde over (y)}0,i={tilde over (y)}i−μi. Interestingly, if
then sign
which is +1 if the variable is positive and −1 if it is negative. Taking this into account, we can rewrite Equation (5.17) by breaking up the domain of {tilde over (y)}0,i:
We should now be able to see that if the input variable is larger than ½ in magnitude, the gradient of the rate loss is constant. This implies that any gradient biases are guaranteed to vanish for noise quantisation proxies when |y0,i|>Δ since the additive noise has a maximum magnitude of ½. This entails the nice equality that
For STE quantisation proxy, the same holds true but for
As justification, R of a Laplacian entropy model and compares it against one of a Gaussian model, where the gradients are biased to the quantisation. It shows rate loss curves (solid curves) and their gradients (dashed curves). Left: Laplacian entropy model. Since the gradient magnitude is constant beyond
the gradient signal would always be equivalent for a rounded latent variable ŷi=└yi┐=yi+ε(yi) as for a noise-added latent if |yi|>Δ. Right: Gaussian entropy model. The same does not apply for a Gaussian entropy model, where it is clear that
One of the unwanted effects of closing the entropy gap, such as with STE quantisation proxies, is that the discretisation of the entropy model inhibits continuity in the gradient flow. Since the probability space is discrete, our gradient signals will also be discrete and dependent on the values of the quantised variables. Unfortunately, this has detrimental effects on the optimisation task.
Consider the dashed plots in
However, the inherent rate-distortion tradeoff prevents a total collapse of the latent space to zero from happening. The distortion requires information in the latent space to be maintained, and so it encourages the latent variables to spread away from zero. The combined effects of STE quantisation ignoring smoothness of values within a quantisation bin and the counteracting gradient signals of the rate and distortion losses yields a phenomenon which has been dubbed the twin tower effect. The results of this is that latent values cluster heavily around the first quantisation boundaries on each side of zero, most often −0.5 and +0.5 for integer quantisation. See
One immediate remedy for this phenomenon would be to penalise latent density from accumulating at quantisation boundaries. This has the effect of introducing auxiliary gradients which are missing from the rate loss when ŷi is zero, and thus assists in moderating the gradient gap. This could be done with a penalty function added to the loss function weighted with a coefficient that yields the maximum value when |yi|=0.5.
Example: We could append our loss formulation from Equation (5.1) with a penalty loss term, λh(y, σ) where λ is a weighting coefficient and
is a penalty loss that is maximal at magnitude 0.5. The extent of the penalty can be adjusted with the σ parameter, which becomes a tunable hyperparameter.
Having visited the effects of STE quantisation, we recognise that the negative impact stem from the entropy gap, or in other words due to the fact that the probability model is discretised. In fact, training with STE achieves much lower reconstruction losses than training with noise quantisation proxies, indicating that the decoder ƒdec seems to benefit from being aware of the true quantised variables.
Since we have distinguished two separate components for the discretisation gap ((−0.5, +0.5) for the rate component of the quantised latents. For the distortion component of the quantised latents, we leverage true quantisation in the forward function and override the backward function with a differentiable straight-through estimator,
We call this quantisation scheme split quantisation. Whilst the discretisation gap remains open for the rate loss, the distortion discretisation gap is effectively closed. On the flip side, this also introduces a gradient gap for {tilde over (Q)}D.
We can address the issue of the new gradient gap for {tilde over (Q)}D by simply rerouting the gradient signal through {tilde over (Q)}R(yi) instead using detaching (or stop-gradient) operations. These exist in automatic differentiation packages which breaks the gradient flow through the detached quantities. With this knowledge, we introduce the soft-split quantisation {tilde over (Q)}SS:
{tilde over (Q)}SS(yi)=detach({tilde over (Q)}D(yi)−{tilde over (Q)}R(yi))+{tilde over (Q)}R(yi) (5.19)
Now, since the gradients are flowing through the rate quantisation proxy, which has a closed gradient gap, we have successfully closed the discretisation gap for the distortion without yielding negative side-effects.
Schematics for both split quantisation and soft-split quantisation can be seen in
Mathematically, the derivative of a true quantisation function is zero almost everywhere and infinity at quantisation boundaries. Hence, this prevents us from using automatic differentiation packages to compute its gradient for further backpropagation. However, since most true quantisation functions can be seen as non-linear operators, we can assign a differentiable neural network ƒQN to simulate the task of the true quantisation function, which we call QuantNet. By supervising it to output the truly quantised variables ŷ=ƒQN(y)≈ŷ, we could leverage its differentiability to backpropagate signals through the quantisation operation.
For each set of latents y, we compute the non-differentiable true quantised latents using, for instance, a rounding function ŷ=└y┐. Then, we supervise the QuantNet with a regularisation term (a norm-distance of degree p, which is the user's input choice) from the ground-truth quantised variablesQN=∥ƒQN(y)−ŷ∥p (5.20)
which we minimise jointly with the standard rate-distortion objective (Equation (5.1)). As a result, the QuantNet is trained to output the true quantised variables which can be used or further propagation through the decoder and entropy model. In order to avoid the network from cheating (for instance by setting the QuantNet to the identity function which would in effect imply no quantisation), the regularisation term has to be appropriately scaled to enforce the intended behaviour of QuantNet.
The architectural details are not specific to this innovation, and can be arbitrarily composed by traditional deep learning operations (linear layers, convolution layers, activation functions, etc.). From the standpoint of the quantisation gaps, QuantNet attempts to narrow the gap of the discretisation gap and entropy gap, and definitely close the gradient gap thanks to its differentiability.
Variations and alternative strategies of QuantNet-based quantisation include, but are not limited to:
The learned gradient mapping approach can be seen as being related to the QuantNet concept. In contrast to parametrising a network to compute the forward function (and its derivative), this approach utilises the chain rule (Equation (5.4)) to parametrise and learn an alternative gradient function
of a true quantisation operation ŷ=Q(y). It can be seen as the generalisation of STE quantisation with a learned overriding function instead of the (fixed) identity function.
A flexible way of learning a gradient mapping is by using a neural network ƒGM:
and optimise over its parameters. If the quantisation gradient
a can be appropriately learned, this innovation contributes to closing the gradient gap for STE quantisation proxies (since in the forward pass, we would be using true quantisation).
There exists at least two possible ways of training ƒGM:
With a continuous relaxation for our probability model, the network spends efforts to optimise for small perturbations in y. From the perspective of the forward function of quantisation, these perturbations yield very little meaning since most of them get rounded away in inference. However, in network training, thanks to our rate formulation (Equations (5.11) and (5.12)), the probability mass evaluated differs by large margin from the actual probability mass assigned in inference, when we actually quantise (see upper two plots in
y (encoder)
[0] ← autoencoder(x, θ[0])
Arbitrary random distribution
Reset encoder weights to
[0])
Backpropagate with ψ[k] which
Gradient descent step for θ
[k] ← autoencoder(x, θ)
[0],
[1], . . . ,
[K]}
Update parameters for fGM
[0])
We can counteract this effect by utilising more “discrete” density models, by soft-discretising the PDF to obtain less “smooth” continuous relaxation, such that the entropy gap can be reduced between training and inference. See the lower two plots in
Since quantisation affects both the rate and distortion terms, it has a major impact on the optimisation task. However, in most cases, we set the bin widths Δ to be constant for all the elements that we quantise. This makes an implicit assumption that every element has the same sensitivity to quantisation errors. We have established that this is unlikely to be the ideal case. For instance, if a certain element y1 is more sensitive to small perturbations than other elements y2, then we would ideally like its error magnitude |ε(y1)|=|Q (y1)−y1|∈ [0, Δ1] to be smaller in general than the error magnitude of the latter |ε(y2)|=|Q(Y2)−y2|∈[0, Δ2]. This is achieved by reducing the bin width of the former with respect to the latter element, Δ1<Δ2.
Inspired by traditional compression whose main source of lossiness stems from the coarseness of the quantisation, we provide a learned method of context-aware quantisation for which Δ is predicted or optimised for. It can be an add-on to (uniform) noise quantisation, for which Δi=1.0 normally for integer quantisation, on top of which we provide the following enhancements:
and the true quantisation function becomes
and then take into account the bin widths in the rate estimation. Optimise for Δi or its precursor during training.
Note that this method does not necessarily aim to close any of the three gaps of quantisation. Rather, its goal is to assist in the parametrisation of the entropy model, of which quantisation is closely linked, to achieve lower bitrates in the compression pipeline.
We have seen multiple cases of where dealing with discrete values encumbers gradient-based optimisation approaches such as gradient descent. However, the task of compression, in particular the inevitable quantisation process within it, is inherently a discrete one. Hence, there is interest in bridging the gap between the discrete and continuous spaces, and one effective way of doing so is through dequantisation. This is the process of producing a continuous distribution of an inherently discrete one (so a kind of quantisation inverse), which can be modelled with a continuous density model.
This concept has strong applicability in areas where a continuous density model (such as our entropy model) is necessary. Different dequantisation methods impose different assumptions of the underlying discrete model. For instance, adding independent uniform noise to dequantise discrete variables imply no assumptions of the dependency between the underlying variables. This is the most naïve form of dequantisation; in reality, for the case of latent variables exhibiting strong spatially local dependencies, quantisation residuals are strongly correlated. Therefore, it makes sense to incorporate more sophisticated dequantisation techniques that can support more realistic continuous probability models.
Some of the dequantisation techniques that we consider in our innovations include, but are not limited to:
Intuitively, one may associate the minimisation of quantisation residuals to having the least adverse effect on the compression optimality. However, due to the highly non-convex nature of neural networks, this is not necessarily true. Rather, we have established theory that the that the minimisation of the adverse effects of quantisation on the global loss definition of compression (Equation (5.1)) is related to minimising second-order effects on the quantisation residuals.
The theory that underpins this assertion can be derived by assuming that the loss term is a function of the input vector x and a feature (latent) vector y, (x, y). Then, given a (discrete) perturbation on the feature vector, Δy, we would like to minimise the following:
[
(x y+Δy)−
(x,y)] (5.21)
Expanding Equation (5.21) using the second-order Taylor series approximation, we obtain
where we can denote for the expected value of the loss gradient vector and Hessian matrix with respect to y:
gy=[∇y
(x,y)] (5.23)
Hy=[∇y2
(x,y)] (5.24)
The loss gradient gy is computable through automatic differentiation packages (through vector-Jacobian product computation). Although the Hessian Hy is also retrievable in the same way, the Hessian is an order of complexity larger than the gradient, and may not be feasible to compute. However, we can often evaluate Hessian-vector (or vector-Hessian) product directly with automatic differentiation tools, circumventing the issue of storing the Hessian matrix explicitly. Nevertheless, we may also use techniques to approximate the Hessian, such as
Example: Assume we have a fully trained network and that a set of unquantised latents y with corresponding data point x minimises a loss function (x, y). Instead of rounding to the nearest integer, we optimise for the quantisation perturbation Δy that has the least impact on the loss value
where the loss gradient is ignored since it theoretically evaluates to zero for a fully trained network. Here, each set of element in Δy are chosen such that each yi is either rounded up or down:
Δyi∈{└yi┘−yi,┌yi┐−yi} (5.26)
This turns the optimisation task defined in Equation (5.25) into a quadratic unconstrained binary optimisation problem; an NP-hard problem, unluckily. However, there exist methods that we could use to approximate the global solution to Equation (5.25), for instance using AdaRound which turns the optimisation into a continuous optimisation problem with soft quantisation variables.
Up until now, we have assumed that the quantisation is performed in feature space, in particular the latent space in which vectors ultimately get encoded and transmitted. However, all forms of quantisation, both introduced as known in the introduction sections as well as our innovations within this field, are similarly applicable in the quantisation of parameter space. This is particularly in the interest of low-bit network architecture, for which low-bit network quantisation plays an important role. It is also useful for innovations involving the transmission of quantised network parameters included in the bitstream, which is employed for instance in finetuning processes. Hence, our advances do not solely pertain to quantisation in feature space, but similarly in parameter space.
In this section, we present the following concepts regarding quantisation, both scalar and vector quantisation considered, in both feature and parameter space for AI-based data compression with details outlined in the referenced sections. All concepts listed below are considered under the context of quantisation in the wider domain of AI-based data compression.
The end-to-end AI-based Compression pipeline usually gets applied to standard format images (single images of arbitrary resolution) and videos (single videos of arbitrary resolution). However, this limits the true potential of the end-to-end based principle that the AI-based Compression pipeline pioneers. Here we describe the usage of AI-based Compression on specific data types and for specific markets and show how, and why, AI-based Compression is ideally suited for exotic data type compression.
Specifically, we will look at the extension of the AI-based Compression pipeline for:
There are numerous exotic data types for which traditional compression approaches (non-end-to-end learned techniques, e.g. JPEG, WebP, HEIC, AVC, HEVC, VVC, AV1) do not work. This is because of three reasons: First, the traditional compression methods have long, costly development processes with a lot of hand-crafting and human input required. Thus, exotic data compression is not a sufficiently big market to justify developing a stand-alone compression approach. Second, the exotic data types often come with a “specific” structure that could be exploited for better compression, but this structure is too complex and challenging to model for the traditional compression approaches. Third, the exotic data often requires different optimisation criteria than compression wrt “pleasing visual images/videos”. E.g. for medical data, the visual aspect is less relevant compared to the accuracy of medical post-processing algorithms that use the compressed data as input. Given these difficulties, until now, there exist only general-purpose compression codecs without specialised sub-codecs. If “sub-codecs” exists, they are often naively-applied traditional methods without any specialisation or modifications.
In contrast to the traditional compression techniques, the AI-based Compression pipeline is an end-to-end learnable coded based on neural networks. As neural networks are universal function approximators, the AI-based Compression can, theoretically, model any dependencies as long as sufficient training data with these structures are given. Also:
Thus, the AI-based Compression is ideally suited to not only create a better general-purpose compression codec but to create numerous sub-codecs that, for the first time, can compress exotic data optimally. In general, these techniques can be extended to arbitrary image and video related exotic data.
In video compression, we have input across the temporal domain, x1,t and x1,t+1, and have temporal constraints on the data: x1,t+1 will depend in x1,t. The temporal constraints are motion-constraints and come from prior knowledge how things move in the time direction.
In stereo data compression, we have two input images or videos, x1,t and x2,t, at the same temporal position. Additionally, if the data's viewpoints overlapped to any degree or have overlapped in the past, the stereo data has spatial constraints given by its common (3D) viewpoint.
Compression can be just another word for redundancy reduction. Stereo data compression aims to use these spatial constraints to lower the entropy/filesize by using the given spatial constraints to reduce information. Thus, it is crucial that AI-based Compression processes and entropy-encodes x1,t and x2,t simultaneously using their joint probability distribution P(x
The AI-based compression pipeline's extension to model stereo data is to give both as input data for the neural network (→early fusion);
In a compression with input x:
In stereo compression with input (x1, x2):
Compression relates to the rate-distortion trade-off: a distortion term that models human perception vision and a rate term that models the filesize.
However, in stereo-image compression, there is a high likelihood that we are not only interested in the visual quality of output {circumflex over (x)}1 and {circumflex over (x)}2, but also care about keeping the integrity of the 3D-viewpoint. For instance, Stereoscopy (in VR) requires more constraints than just the visual quality of {circumflex over (x)}1 and {circumflex over (x)}2. Stereoscopy also requires that the re-projection of the 2d-plan into the 3d-world is consistent, aka. that the depth information encoded between {circumflex over (x)}1 and {circumflex over (x)}2 is (sufficiently) accurate.
To model this constraint, the AI-based Compression pipeline requires and additional (differentiable) loss term;
Another observation is that with a stereo-camera setting, we have a new type of metainformation that can be helpful: The positional location information of the cameras/images and their absolute/relative configuration. For instance, in Stereoscopy the baseline, the FoV (field-of-view), sensor-type, resolution, and other metainformation is invaluable.
For stereo data compression, we can either have our neural network encode this information as a prior through the training process. Or we can explicitly feed this information into our neural network through architecture modification. For instance, we can bring the metadata to the same resolution as the image/video data through processing with fully connected layers and then concatenate it to the inputs (x1, x2) and to the bottleneck latent y.
If we use early fusion in the compression pipeline, it also makes sense to provide the ground-truth dependencies between x1 and x2 as additional input. This is crucial as a more proper modelling of the joint-probability function of x1 and x2 is a conditional joint distribution given the spatial restrictions and dependencies. E.g. using P(x
Multi-view data is the natural extension of stereo-data to a more complex setting. Instead of x1,t and x2,t, we now have N input images/video {x1,t, x2,t, . . . , xN,t} that might come from different devices with different settings. Good examples include 2D-to-3D, Photogrammetry, Self-driving cars, SLAM-applications, (online) visual odometry, 360°-video (can be interpreted as multi-view data), n-D simulation, 360°-images/videos on a website, panoramic images/videos, and others.
Each of these examples has its unique constraints, which if we can phrase as a differentiable loss function, we can incorporate in the AI-based Compression pipeline. Suppose we can not express the constraints into a differentiable loss function. In that case, we can still use either a proxy network similar to “QuantNet” or “VMAF Proxy” and/or can use reinforcement learning techniques to model it.
If we have data that does not come from the same sensors, e.g. in self-driving cars, the camera sensors around the vehicle can vary; it makes sense to use pre-processing layers to bring all data into the same feature space.
In stereo data compression, we have spatial constraints, usually given by a common 3-D viewpoint. In video compression, we have temporal constraints, traditionally given by motion-vectors and motion/acceleration/momentum restrictions. In multi-view data (especially multi-view video), we tend to have both of these constraints.
Suppose we have a self-driving car with eight cameras at the top-left, top, top-right, middle-left, middle-right, bottom-left, bottom-middle, bottom-right. At the same time, t, some of these cameras will have overlapping viewpoints from a common 3D-scene and thus, have spatial constraints. However, some of the cameras will not have a common 3D-scene at t (e.g. top-right and bottom-right).
If we know the approximate rate at which different temporal data leads to spatial constraints, we can use this prior as helpful information in the AI-based Compression pipeline. There are three ways to do so: First, we use x2,(t+i) and x2,(t+i) as input, if we know that the past frame at t from camera 1 spatially constraints the current frame at (t+i) from camera 2. Second, we include multiple spatial and temporal inputs and indicate video meta-information, which inputs tie to each other spatially. Finally, we can keep a queue of bottleneck-layers {ŷ1,t, . . . , ŷn,t; ŷ1,t+1, . . . , ŷn,t+1; . . . ; ŷ1,t+i, . . . , ŷn,t+i} and model which inputs tie to each other spatially on the entropy-level.
Note that
Satellite image and video sensors usually capture more spectral bands than just the visual spectrum. Thus, instead of the standard format of a 3-channel image/video, we get n-channel image/video with each channel representing a particular band (different wavelengths).
This data type can be seen as a particular case of multi-view data. The viewpoints all overlap, but we get different information about the scene, like having various camera sensors with variable calibration data. Thus, all previously mentioned methods for stereo-data and multi-view data apply, too.
However, in addition to the previous cases, satellite data often comes with an objective which is not primarily scene quality but there are classification or segmentation questions. For instance: Atmosphere forecasting and monitoring, weather predictions, event classification, detection of geological or human events, agriculture monitoring, military purposes, and many others. This provides the opportunity to compress spectral data with a non-visual based loss term for distortion but an event-driven loss based on the exact question that we want to answer. Also, this is not limited to one auxiliary loss but can easily extend to n-auxiliary losses.
Suppose we have satellite data and want to monitor a forest's health and detect oil spills (dual objective). Instead of only having the visual quality as the distortion term, we can model the data-post processing methods' accuracy. Let's assume we have a spectral-image/video algorithm to detect oil spills O(x) and an algorithm to monitor forest health F(x) with some distortion metric D(·). Our satellite data compression objective for input data x with n-channels becomes:
Using this approach in combination with AI-based Compression, we can quickly and cheaply design numerous neural networks specialised in compressing Satellite data with given objectives. Even better, we can switch between different approaches easily. Let's assume we have a compression codec trained on forest-data monitoring, but after a week, we want to reprogram our Satellite for oil spill monitoring. In AI-based Compression, having a codec specialised means having the network trained on specific loss-terms and using the trained network parameters θ for inference. If we want to change our objective, we have to re-train another network, get another 0 optimised for our new goal, and replace the neural network weights. This can all be done in software, with no hardware replacements and can be seen as “streaming a codec”—an invaluable property of AI-based Compression.
Compression with medical data follows the same guidelines as mentioned in the Satellite/Space data section. Often, we have special image and video data (Saitellite==multi-band data; Medical: health-care scans), which will be the input for post-processing algorithms. Thus, we require training over a particular input data training set, and the compression objective needs to be updated (→new θ).
The auxiliary loss terms can be, amongst others, related to:
Learnt image and video compression is based on neural networks which are non-invertible transformations. We provide multiple ways of integrating the standard learnt pipeline with Invertible Neural Networks (INNs), also known as Normalising Flows, which are bijective mappings that can be used to transform a random variable to an alternative representation and back. The bijective ability of INNs allows us greater flexibility in enforcing a prior distribution on the latent space, in addition to providing a point of contact between adversarial and Maximum Likelihood Estimation (MLE) training.
Let us consider a random variable y∈N. We can transform this variable using a mapping ƒ:
N→
N (it is important that the input space has the same dimensionality as the output space). The probability distribution of the output variable z can then be calculated in terms of the probability distribution on the input variable y, as shown below:
This formula has two requirements:
must be defined, in other words the Jacobian matrix has to be square. This has important implications because it means that the normalising flow can't change the dimensionality of the input.
In order to satisfy both of these conditions, and thus to be able to calculate the probability distribution of z in terms of the probability distribution of y, the transformation ƒ needs to be invertible, hence the need for an invertible neural network (although in order to be defined as a normalising flow, the transformation is not required to contain a neural network).
Why do we want to define the output distribution in terms of the input distribution?
Because our objective is to map a complex distribution (such as the distribution of the latent space of our autoencoder when there is no entropy model enforced) to a simple distribution (such as a normal or laplacian prior), so that we can enforce an entropy model on z while retaining the spatial information in y that improves reconstruction quality.
In addition to the hard requirements listed above, there is one more “soft” requirement, that is, the determinant of the Jacobian matrix has to be easy to compute. This is not the case for any matrix, so the calculation can become expensive when the Jacobian matrix has a high rank, especially if we chain multiple transformations together in a normalising flow (a feat that is quite common to increase representational power).
How can we make the determinant of the Jacobian easy to compute?
If the square matrix is upper- or lower-triangular, i.e. it has non-zero elements as shown, for example, in
We remind the reader that the Jacobian matrix of a mapping ƒ: N→
N has the form shown in
I.e. it is an N×N matrix for an input x containing elements {x1, x2, . . . xN} and an output containing elements {ƒ1, ƒ2, . . . ƒN}.
For simplicity, we will describe only the process by which to make this matrix lower triangular, because making it upper triangular consists in a very similar process.
We introduce the concept of a coupling transformation that splits the input x into two partitions xa and xb. The output of the coupling transformation is then:
za=xa
zb=g(xb,m(xa))(7.2)
Where g and m are arbitrary functions.
In an additive coupling transformation, we define g as the arithmetic addition, and m as a neural network. This results in the below transformation:
za=xa
zb=m(xa)+xb (7.3)
This transformation is both trivial to invert, and has a triangular Jacobian.
If we want to retrieve x from z, we only need to apply the following operations:
xa=za
xb=−m(Za)+zb (7.4)
It is important to note that the neural network transformation needs not be inverted. This greatly simplifies the process, as standard neural network architectures are hard to invert.
Additionally, the form of the Jacobian is shown below:
Not only this Jacobian is lower triangular, but its diagonal entries are all 1. Hence, the determinant of the Jacobian is 1, meaning that the computational cost of calculating it is O(1) for any kind of additive coupling.
It should be noted that additive coupling is not the only invertible transformation: there are such things as multiplicative or affine coupling layers, where the mapping g is the element-wise multiplication operation, and the joint multiplication and addition operation respectively.
The additive coupling transformation is said to be a volume-preserving transformation. This stems from the fact that the determinant is 1. Volume-preserving (VP) transformations generally have lower transformational power than non-volume-preserving (NVP) ones, since the formers are prevented from making some eigenvalues bigger and others smaller, resulting in more variation with respect to some input elements than others.
Multiplicative and affine coupling transformations are non-volume preserving. Let us consider an affine coupling:
Za=Xa
Zb=Xb⊙S(Xa)+m(Xa) (7.6)
Where s is another arbitrary transformation that in practice is defined as a neural network.
The Jacobian of an affine layer is below:
Now, since the diagonal entries of this Jacobian are not all ones, the determinant of the matrix is not 1 anymore, instead being the product of the diagonal elements of the scaling transformation s.
Here we describe two additional operations that we use in our normalising flows.
The first operation is the squeeze layer. Although we previously stated that the dimensionality of the input of a normalising flow cannot be different from the dimensionality of the output, we can change the spatial resolution and the number of channels of the feature maps, provided that the total number of elements is unchanged. The change in dimensions is actuated with a squeeze layer that changes the dimensionality of the input tensor RH×W×C into RH/2×W/2×4C. This allows the convolutional layers in the neural networks inside the coupling transformations to operate on different scales at different stages in the normalising flow.
The squeezing operation reallocates pixels from the spatial dimensions into different channels using a checkerboard pattern.
The checkerboard pattern ensures that pixels that are spatially close are allocated to pixels in different channels that have the same spatial location. This mitigates the distortion of spatial information, which is important for the convolutional layers.
The second operation we describe is the factor-out layer.
The factor-out operation splits the feature map in two parts along the channel dimension (although it can also split the feature map in the spatial dimensions). Then, one part is passed as the input to the next normalising flow block, while the other part is passed directly to the final output of the pipeline.
This has two implications: firstly it reduces the computation that needs to be done, which can add up to a great amount given that normalising flows must maintain the dimensionality of the input; and secondly, the loss function is distributed through the whole network, so gradients are directly passed from the loss to all blocks of the chain instead of the last block only.
Finally, there is a trick that we can use to circumvent the limitation of not being able to change dimensionality: if we need to increase the dimensionality of the input, we can pad it with zero values, thus increasing the dimensionality from H×W×C to H×W×C+D. The normalising flow will then produce as output of the same dimensionality as the padded input, that is H×W×C+D: note that this is larger than the actual input size of H×W×C. We just showed this trick in the channel dimension, but we can just as easily apply it in the spatial dimension as well.
As the name suggests, FlowGAN is a generative model obtained by combining a normalising flow and a GAN setup.
In a GAN setup there is a generator and a discriminator network. The input to the generator z is sampled from a prior distribution and the generator network transforms it to the underlying distribution of the training data, i.e. if the training data is natural images the generator learns a mapping from P(z) to the distribution of pixel colours in natural images, that we define as P(G(z)).
The discriminator network is a classifier trained on both the generated images and the real training images: its aim is to differentiate between the training images and the images generated by the generator. Intuitively, when the discriminator is unable to properly classify the two classes of images, then the generator is outputting realistic images that look like the training images.
This adversarial training strategy presents a problem: the losses of the two networks have poor interpretability. For example, there is no correlation between the loss of the generator and how realistic the generated images look, often requiring the addition of visual quality scores to estimate it. This poor interpretability stems from the absence of an explicit likelihood function in the density model, which in the case of GANs is implicit (it has no clearly-defined density model for the generated data y). Unfortunately, it is impossible to obtain an explicit model of P(G(z)), because its density can't be defined in terms of P(z).
FlowGAN solves this problem by using a normalising flow as the generator network. An example structure is illustrated in
Such a setup allows to train the generator in two ways: either with adversarial training, against the discriminator network; or directly with Maximum Likelihood Estimation, using the change of variable formula in Equation (7.1).
Given discriminator losses ho (on the training images) and h′ϕ (on the images created by the generator), the complete training objective of FlowGAN is below:
The first two terms in the equation above are the adversarial terms for “real” and “fake” images, that need to be maximised with respect to the discriminator weights ϕ and minimised w.r.t. the generator weights θ (we remind the reader that the generator needs to fool the discriminator into classifying the fake images as real). The third term is the likelihood of the training data, which is normally intractable in an adversarial setup; however, since the generator is an INN, we can express p(y) in terms of p(z) which has a tractable analytical distribution, so the likelihood is tractable in FlowGAN. The scaler A determines whether training is joint adversarial and MLE, or only adversarial (in the case where A is zero).
In this subsection we describe the substitution of both the encoder and decoder networks with a single INN. The types of normalising flow layers we use include (but are not limited to):
Using an INN instead of two networks results in an overall smaller size on computer disk which translates to faster transfer of the compression pipeline over communication networks, facilitating the transfer of fine-tuned codecs for specific sets of images or videos on-the-fly.
This pipeline is valid with a continuous flow, but it can be used with discrete flows as well with a small modification. In the case where a continuous flow is used, the quantisation operation is necessary in order to obtain a quantised latent space that can then be arithmetic coded.
Now let us consider the case where a discrete flow is used. Discrete flows have a similar structure to continuous flows. For reference, we illustrate the architecture of the Integer Discrete Flow in
The peculiarity of a discrete normalising flow is in its coupling layers, where a quantisation operation is included (see below).
xa=Za
xb=└−m(za)┐+zb (7.9)
Assuming a discrete input x, the output z will also be discrete, since the only possible source of non-discretised output is the neural network transformation m, and that is explicitly quantised in the coupling layer above. Thus, the quantisation operation described in the pipeline illustrated in
We also remind the reader that images are stored on computer disk using integer values, so they are already quantised before processing. This changes the compression pipeline from lossy to lossless, as the latent space can be arithmetic coded as it is, because it is already discrete when it is output by the integer normalising flow.
In addition to the pipelines described in the previous subsection, we provide the addition of a normalising flow to the encoder-decoder pipeline to help model the prior distribution. In particular, we provide the substitution of the hyperprior network with a normalising flow.
The additional variable w decouples y from the entropy model, resulting in better reconstruction fidelity. Additionally, the encoder and decoder downsample the input, acting as a preprocessing step that is complementary to the INN transformation that is unable to down- or up-sample. We highlight the necessity of a discrete normalising flow in this pipeline, since the input ŷ is already quantised, and w is also required to be discrete since it is directly fed to the arithmetic coder. An example is shown in
In this subsection we describe a modification we make to normalising flows using hyperprior networks. This modification has wide applications, being virtually useful for any application of normalising flows in image compression. The advantage they provide is a further reduction in bitstream size, due to the application of a hyperprior network at each factor-out layer inside a normalising flow.
Normally, the factor-out layers of a normalising flow are connected to a parameterisation of the prior distribution modelled with a neural network.
In this integration, illustrated for example in
Both the outputs of a factor-out layer (y and z) are fed to a hyperprior autoencoder that outputs a feature map b. This feature map is then concatenated to y and fed to the parameter estimation model. By compressing y and z into w and sending this latent as side-information, we further improve the compression ability of the pipeline.
Mutual information (MI) between X and Y is defined as follows:
We model p(y) and p(y|x) using INNs. This gives us approximations to these values, which we can use to compute the I(X; Y). By being able evaluate I(X; Y), we are able to maximize the mutual information between X and Y. A use case for this method is further described in our Mutual information section, Section 8.
In this subsection we introduce a form of meta-compression that sends model weights along with image data. Sending model-specific information enables a more flexible compression scheme, where the coding system is tuned to the compressed image.
The neural weights of the network need to be compressed in order to be sent as bitstream, in a similar way to how the latent space of the autoencoder is compressed. This requires two things:
Generally, the distribution of the weights of a neural network can be very complex. Hence, this is a suitable application for normalising flows; not only that, we can quantise the weights readily with minimal loss of performance by taking advantage of Quantisation Aware Training (QAT), a standard feature of deep learning frameworks such as PyTorch and TensorFlow. After quantisation, we pass the weights of the neural network to an INN, that returns a processed representation of the weights following the prior distribution, and then we can encode this representation with an arithmetic encoder. The decoding process comprises using an arithmetic decoder to retrieve the processed weights, and then undo the transformation by passing them to the inverse of the normalising flow. An example illustration of such a pipeline is shown in
Below we enumerate the concepts which relate to the present section.
In the field of learnt image and video compression, an aim is to produce the most efficient encoding of an image, meaning that there are no redundancies that are not required by the decoder to re-produce an accurate compressed output. Another way this can be viewed is the requirement that the encoder must discard redundant information; this is implicitly happening, however a way to further improve compression efficiency is to explicitly model the inherent dependency between the input and output, known as mutual information. Here mutual information for learnt compression is discussed and novel methods of training compression pipelines with this metric are given.
Mutual information is a quite esoteric yet fundamental property that is useful to represent relationships between random variables. It cannot be well understood without a background in statistics and probability theory. For two random variables the mutual information I(X; Y) may be most intuitively expressed in terms of entropies H:
I(X;Y)=H(X)−H(X|Y) (8.1)
This means that the mutual information between X and Y is equal to the reduction in the uncertainty (or entropy) of X (H(X)) reduced by how much we know about X if we are given Y (H(X|Y)). H(X|Y) is a measure of what Y does not tell us about X, i.e. the amount of uncertainty that remains about X after Y is known. The equation above can thus be reformulated in terms of text as follows:
The amount of uncertainty in X, minus the amount of uncertainty in X that remains after Y is known, which is equivalent to the amount of uncertainty in X that is removed by knowing Y.
This is shown in
Mutual information may also be expressed for probability density functions as follows:
This is equivalent to the KL divergence between a joint distribution p(x, y) and marginal distributions p(x), p(y). Mutual information can be more succinctly expressed in terms of this F divergence:
I(X; Y)=DKL(p(x, y)∥p(x)⊗p(y)) since the joint p(x, y) may be expressed as a conditional p(x, y)=p(x|y)p(y). The intuitive meaning of mutual information expressed with the KL divergence is that the larger the divergence between the joint and the product of marginals, the stronger the dependence between X and Y.
Mutual information has found a wide range of meaningful applications within the field of data science.
For our field of image compression, it should be clear that there is a notion of mutual information between an input image and its corresponding compressed output from our compression pipeline. In fact, it turns out that for an autoencoder that is trained with an MSE loss, the network is learning to maximise (a lower bound on) the mutual information between the input and the latent representation Y, I(X; Y). This should make intuitive sense because by maximising I(X; Y) we are compressing away information in Y that is not necessary to retrieve {circumflex over (X)} from Y, hence why a strong correlation between the input X and Y is expected, which is what we observe for the latents (the compressed bitstream) of our models.
Notwithstanding the usefulness of mutual information for a wide range of fields, estimation of mutual information for unknown probability densities still remains intractable to compute. It's only tractable for discrete variables, or for a limited set of cases where known probability distributions may be applied.
Therefore there have been a number of efforts to provide estimators that can provide a tight lower or upper bound on the mutual information. In this section, the Barber & Agakov and InfoNCE bound is defined.
The Barber & Agakov upper bound is defined as follows:
For the case of the lower bound on the mutual information, we replace the intractable conditional distribution p(x|y) with a tractable problem over a variational distribution q(x|y):
The Barber & Agarov lower bound is tight when q(x|y)=p(x|y).
InfoNCE is a lower bound where a critic may be used to estimate the joint and marginals.
The role of the critic is to learn to predict p(y|x). The critic may be parameterized by a neural network.
In this section, a novel approach that seeks to maximize the mutual information between the reconstructed output of the compression pipeline and the input is explored. As explained in the introduction, maximizing the mutual information is a method of producing a tight coupling between two parameters.
In general, it is not clear how well mutual information estimators work for high dimensional problems, such as image compression and the extent of the lower and upper bounds is hard to define. The estimators may be biased or provide inaccurate estimates. A novel way around this issue is to treat the compression pipeline as a simple channel, the aim is to increase the channel capacity—increasing channel capacity equates to increasing the amount of information that can flow through the corrupted channel. The highest channel capacity is achieved when the noise added by our channel is zero. When the input x passes through the channel it is corrupted by noise n, as shown in Equation (8.6). Our aim is to maximize the channel capacity by maximizing the mutual information between the input x and the output z, essentially learning to remove corruptions introduced by the noisy channel.
{circumflex over (x)}=x+n (8.6)
Modelling the input x and the noise n as zero-mean independent Gaussian tensors, (0, σ2) it is possible to compute a closed-form solution of the mutual information I(x; {circumflex over (x)}) where x is the input and {circumflex over (x)} is the compressed media.
The parameters σx and σn of Equation (8.7) are learnt by neural networks. In terms of entropy modelling, the MVND entropy model may be used to model our source x and our noise n. However, in general, any type of density estimation approach (such as Maximum likelihood or Maximum a posteriori) as well as any generative model (such as a PixelCNN, Normalizing flow, Variation auto encoders) can be used. The aim of the training is to force our encoder-decoder compression pipeline to maximise the mutual information between x and {circumflex over (x)}, which forces our output to share information with our ground truth image. The training may be executed in multiple different ways.
The first method of training is to directly maximise mutual information in a one-step training process, where the x and n are fed into their respective probability networks S and N. The mutual information over the entire pipeline is maximised jointly. This is shown in (x,{circumflex over (x)})=R(x)+λD(x,{circumflex over (x)})+αI(x;{circumflex over (x)}) (8.8)
The second approach is a bi-level or two-step process. Firstly, the network S and N is trained using negative log-likelihood to learn a useful representation of σn, and σx, based on the closed-form solution of the distribution selected. This part of the process is shown in
In general, for any function ƒ and g it holds that I(X; Y)≥I(g(X); ƒ(Y)) where I(X; Y)=I(g(X); ƒ(Y)) if and only if ƒ and g are invertible and volume-preserving, i.e. det(ƒ)=1 and det(g)=1. As such the noise n and/or the input x can be transformed by an arbitrary function as long as the constraints above apply, e.g. ƒ and g could be an invertible neural networks (INN). The invertible transformation can be applied to either X or Y or both. A particular analytical example of ƒ and g could be an orthogonal basis transform into another basis, or converted into another domain, such as the wavelet domain, to better model the probability distributions.
In addition, the approach may also be applied on patches or segments of images. Also a multi-scale approach may be used, this is naturally the case when the transformation above provides multiple different scales, such as the case given the wavelet transform, where mutual information for each scale is computed and then aggregated. This approach may also be further generalised to a multivariate distribution where the tensor to be modelled is split into blocks (in spatial and or channel dimensions) of variable sizes and modeled using a multivariate normal distribution with a mean vector and co-variance matrix per block of elements.
Finally the distribution used to model the source and noise is not limited to a multi-variate Gaussian distribution, but may be extended to any continuous distribution such as Behrens-Fisher distribution, Cauchy distribution, Chernoff's distribution, Exponentially modified Gaussian distribution, Fisher-Tippett, log-Weibull distribution, Fisher's z-distribution, skewed generalized t distribution, generalized logistic distribution, generalized normal distribution, geometric stable distribution, Gumbel distribution, Holtsmark distribution, hyperbolic distribution, hyperbolic secant distribution, Johnson SU distribution, Landau distribution, Laplace distribution, Levy skew alpha-stable distribution or stable distribution, Linnik distribution, logistic distribution, map-Airy distribution, etc,. This allows for more accurate modelling of the source and noise while maintaining a close formed solution.
A novel method of performing compression using mutual information, that does not involve a noisy channel, is to explicitly optimise the mutual information of the output and the input of the neural network such that this metric is maximised. The mutual information estimator used is not restricted to the bounds presented in the earlier sections such as Barber & Agakov or InfoNCE, along with bounds not presented explicitly, such as TUBA, Nguyen-Wainwright-Jordan (NWJ), Jensen-Shannon (JS), TNCE, BA, MBU, Donsker-Varadhan (DV), IWHVI, SIVI, IWAE, etc., Moreover, neural networks (non-limiting examples include: INN, auto-encoders, conditional model) can also be applied to estimate probability estimates p (x, y), p (x|y) for mutual information estimates. The loss of the neural network is therefore augmented in the following way:(x,{circumflex over (x)})=R(x)+λD(x,{circumflex over (x)})+αI(x;{circumflex over (x)}) (8.9)
where x is the input, {circumflex over (x)} the output, R the estimated rate, D compression distortion, I the estimated mutual information. λ and α are scaling coefficients. A simplified generic example of the compression network can be seen in
An extension of the mutual information, defined in Equation (8.2), appropriate for video content or temporally correlated media is to condition the joint and the marginals based on N past data points, c. The conditioning may be applied using the compressed output ci={circumflex over (x)}i or the ground truth input ci=xi. Conditioning on the compressed media allows for a temporal reduction of artifacts by enforcing logical and temporally consistent consecutive frameS.
During variational approximations, the conditional or marginal may be parameterised as a neural network; the conditional approximation would be given as input the previous N samples, in addition to the current ith sample.
In the previous sections, mutual information optimisations was performed for the input and output of the compressed media, by computing I(x; {circumflex over (x)}), however, this can be extended to optimise for bit-rate R. Maximising mutual information of the latent parameter y and a particular distribution , as seen in Equation (8.11), can be used to optimise for rate.
This is because rate is computed using Equation (8.12), where qy is the tractable probability distribution estimate of py by an entropy model.
R=H(py,qy)=Ey˜p
When I(y; n) is maximised, py≈, such that the unknown distribution py can be modelled as a noisy known distribution
, this provides a more efficient entropy computation. The mutual information of I(y; n), as shown in Equation (8.13), requires that py and
be dependent. In simplified form
which is maximised if the joint probability is large and the marginal probability is small, i.e. strong dependence.
For the case where has a known closed-form solution, this can be optimised for directly using negative log-likelihood, as shown in for example
=N(0, 1), the mutual information can be computed using Equation (8.7).
In the example of =−L.
In learnt image and video compression, the latent space is normally conditioned to follow a certain distribution using maximum likelihood estimation (MLE). We describe alternative approaches to learnt compression that integrate and exploit other methods of enforcing specific densities on latent spaces. This allows us to circumvent some limitations of MLE and obtain greater flexibility in the classes of distributions that can be modelled.
Learnt image and video compression mainly consists in three components: an encoder neural network, an entropy model, and a decoder neural network (the encoder and decoder networks together are referred to as an auto-encoder). The encoder network processes the image or video into a representation called a latent space, and the decoder network applies the reverse transformation. The encoder network applies a first pre-processing step before the entropy model is applied.
The entropy model is represented as a uni- or multi-variate probability distribution pm (y), normally assumed to have a parametric form (for example, a standard Normal distribution, or a Laplacian distribution, etc). The parameters of the entropy model are usually fitted to the training data (using methods like maximum likelihood), although this is not a requirement—it only improves the compression efficiency of the pipeline. On the other hand, the actual marginal distribution of the data p(y) is not known in advance.
With an entropy model in place, we can further compress the latent space using an entropy code such as Huffman coding or arithmetic coding: the amount of bits B contained in this code (which is the bitstream that is used as the final compressed representation) can be calculated using Shannon's cross-entropy.
B=Σp(y)log2(Pm(y)) (9.1)
This quantity is minimised when p(y) and pm(y) are the same probability distribution, i.e. when the distribution in the entropy model matches the real distribution of the latent space.
Fortunately, we can directly train our models towards this objective, in fact the usual form of the loss function of a learnt compression model is
L=D(x,{circumflex over (x)})+λB(y) (9.2)
Where D is a distortion loss between the original image and the compressed+decompressed image (less distortion equals better fidelity.
An important concept to keep in mind is that Shannon entropy is only valid on discrete sets of symbols. This means that, in order to apply arithmetic coding on the numerical values inside the latent space, we need to quantise these values.
Quantisation is a big problem in learnt image and video compression, because the quantisation operation has no gradient, and our networks are trained using gradient descent, so it requires all operations inside the pipeline to be differentiable. In practice, a differentiable substitute is used instead of quantisation during training, for example the addition of noise, or the Straight-Through Estimator (STE); however, this is just an approximation of the real operation.
What if we could bypass the quantisation operation?
Here we address this question.
GANs are often used when it is not clear what form the loss function should have. This is especially applicable to generation tasks, where the loss needs to define what a realistic-looking image looks like. For instance, if a model is trained to generate human faces, the loss should contain information such as what is a realistic-looking nose, a realistic location for eyes and mouth, a realistic skin color, etc. Such a loss is impossible to craft manually, hence we substitute the loss function with a second neural network referred to as a Discriminator.
The discriminator is trained as a classifier that needs to differentiate between the images generated by the generator, and the images in the training dataset. The generator network has the opposite objective: to generate images that will be classified as real, despite being generated by the network. In artificial intelligence, this is referred to as a zero-sum minimax game: zero-sum because the loss of the generator is directly opposite to the loss of the discriminator; and minimax because the objective of the networks is to minimise the loss of each network in the worst possible case, that is when the loss of the other network is at a minimum.
As described in a previous subsection, learnt compression pipelines make use of an entropy model to further compress data. This is done with maximum likelihood estimation on the latent space, under the prior assumption that the probabilities of its values follow a certain distribution. Adversarial training can be an effective alternative to maximum likelihood: indeed, AAEs make use of GAN-style training to enforce a specific distribution on their latent space.
The task of the discriminator network is to differentiate between the latent space and samples from a known prior distribution. Conversely, the task of the generator network (in AAEs, the generator is the encoder) is to generate latent spaces that are indistinguishable from the prior distribution.
AAEs are autoencoders that use GAN-style training.
The biggest advantage of AAEs as opposed to autoencoders trained with MLE is that AAE training is sample-based, while MLE requires parametric distributions with analytical form. Examples of parametric distributions include normal, laplacian, beta distributions, etc. This puts a strict limit on what class of distribution the latent space is allowed to follow, because many distributions have no analytical form but can be useful priors (for example, categorical distributions where the values can only assume one of a finite set of values).
Analytical distributions have a density that can be represented as a formula. For example, a Normal distribution has a probability density function defined as
This means that the density can be simply calculated at any point in the distribution's support (the set of numbers it is defined on). Having an analytical form also enables a range of techniques such as differentiable sampling in the form of the reparametrisation approach (we remind the reader that operations used in a learnt pipeline need to be differentiable in order to work).
On the other hand, an example of a non-analytical distributions is a categorical distribution where the only information we have is a few samples as listed below:
l={0,0,3,1,0,2,3,0,2} (9.4)
We cannot backpropagate through such a distribution, hence it is more problematic to include in a learnt compression pipeline.
It is useful to understand how the difference between one probability distribution and the other can be calculated, as there are innumerable methods for estimating distance between distributions.
KL Divergence: a widely used method in machine learning is the Kullback-Leibler (KL) divergence, sometimes referred to as relative entropy. The KL divergence between distributions P and Q is defined as
Where p(x) and q(x) are the densities of the distributions at point x.
This distance has a limitation. As shown in (9.5), the density p(x) at all points needs to be known, and this is only the case when the distribution has an analytical form. So, the KL divergence can't be used for all distributions.
Moment Matching: a simpler way of comparing distributions is simply to calculate their moments and compare the corresponding moment of one distribution against the moment of the other. Moments of a distributions are numbers that describe the shape of its density, for example the first moment of a distribution is its mean, the second moment is its variance, the third moment is the skewness, etc. In order to quickly calculate a measure of difference between two distributions, we can calculate the difference between the mean of one and the mean of the other, then the difference between the variance of one and the other, etc. This has the advantage of being completely sample-based, that is, we don't need to know the analytical form of the distribution, we just need to be able to draw samples from it.
MMD: maximum mean discrepancy is another method that does not require the distributions to have an analytical form. It is weakly related to moment matching, and could be considered a generalisation of it.
Let us define a kernel h that maps from the set to the set
. Maximum mean discrepancy is then defined as
MMD(P,Q)=∥x˜p[h(X)]−
y˜Q[h(Y)]∥
(9.6)
That is, the norm of the difference between the expected value of the kernel embedding of the first distribution and the second. As a simple example, if we pick h to be the identity function, MMD reduces to first moment matching (i.e. the embeddings collapse into the mean of the distributions).
Optimal Transport: this family of methods stems from the field of operations research. The distance between distributions is formulated in terms of finding the most efficient transportation plan that moves probability mass from one distribution to the other.
The most well-known measures of distance in optimal transport theory are Wasserstein distances.
Mathematically, if we define a transportation plan between distributions as γ from the set containing all transport plans Γ. and a cost of transport c, W-distances are defined as below:
That is, the minimum-cost transportation plan to move all the mass of P into Q. The transportation cost is usually the L1 norm, in which case the distance is the Wasserstein-1 distance.
Just as MMD, Wasserstein metrics are also purely sample-based, i.e. they can be used with any probability distribution regardless of whether they have an analytical form. However, W-distances are non-trivial to compute because they require finding the minimum-cost transportation plan. This is an optimisation problem, which is non-differentiable and can be extremely computationally intensive when the distributions are very high-dimensional.
Sinkhorn Divergences
Sinkhorn divergences can be considered a generalisation of both MMD and Wasserstein metrics. Mathematically they are formulated as below:
Where W∈ is a regularised form of Wasserstein distances defined as below:
If we compare against Equation (9.7) we can see that an additional KL divergence term has been added. This has the effect of mitigating the main problem with Wasserstein distances (their non-smoothness and subsequent computational expensiveness).
We present three general frameworks where the latent space of an auto-encoder is forced to follow a particular distribution by a joint training process.
The algorithm of the first framework is detailed below:
H×W×C
This framework is equivalent to the standard learnt compression pipeline, where the prior distribution is embedded inside the entropy loss. The difference in our approach is the choice of LB: while the standard pipeline uses KL divergence, our choice is more free in that we also use moment matching as one of the divergence measures, which has not been done previously.
The algorithm for the second framework is below:
H×W×C
The above algorithm describe the adversarial auto-encoder setup that we use for image compression. This allows us to force the latent to follow a sample-based distribution that has no analytical form.
In addition, we use a variety of adversarial setups for the generator and discriminator. The first category is class probability estimation, which includes all losses in
The second category is direct divergence minimisation using f-divergences such as:
The third category is the direct minimisation of a Bregman divergence, and the fourth category is moment-matching.
The algorithm for the third framework is below:
H×W×C
This framework is easier to train than framework 2 because there is no adversarial training. Additionally, it is more flexible than framework 1, in that the entropy loss calculation depends purely on sampling from the prior distribution and comparing the sample against the latent space using one of the following measures:
Mean Maximum Discrepancy is differentiable, and so are Sinkhorn Divergences. But pure Optimal Transport measures are not. A contribution of ours is a simplification of Wasserstein distances that exploits the fact that W-distances are differentiable in the special case where the distributions are univariate.
In the 1-dimensional case, Wasserstein collapses to the following definition:
Where M is the number of elements in p and y, which are the sample from the prior distribution and the latent space respectively. Indices i[m] and j[m] are the indices of the sorted values of the tensors in ascending order.
As we can see in the equation above, we no longer need to find the infimum of the transport plans, so the optimisation problem is done away with completely. An illustration of what this divergence measures is shown in
With univariate distributions, calculating the Wasserstein-1 distance is equivalent to calculating the L1 norm between the sample and the latent, once their elements have been sorted by value. This results in a simple code implementation, defined below:
N
N
Naturally, the algorithm outlined in Algorithm 9.4 is not limited to use the Wasserstein-1 distance: for example, if calculating the Wasserstein-2 distance is required, all that is needed is to substitute the L1 norm with an L2 norm.
A limitation with Algorithm 9.4 is that the input tensors are flattened before sorting, thus it only supports univariate distributions, which have a much smaller representational power compared to multivariate distributions. However, we circumvent this limitation by defining a separate prior distribution for each channel or pixel in the latent space, then sampling from each of these distributions (see
Note that in
After sampling the target tensor, we calculate the Wasserstein distance separately for each pair of corresponding pixels in the latent space and sampled tensor, as below:
Wu,v=W1D(Su,v,Yu,v) (9.11)
And finally we aggregate all these individual Wasserstein distances by averaging:
Note that when using W-1 distance channel- or pixel-wise, a large batch of images is required to obtain a large enough sample size.
Using sample-based entropy losses, such as what is used in framework 3, unlocks a new capability with our models, that is, enforcing a discrete distribution on the latent space without explicitly using the quantisation operation.
The training pipeline is completely unchanged from the one associated with framework 3, the only difference being that the prior distribution is now a discrete distribution instead of being continuous.
The absence of an explicit quantisation operation means that, during training, the encoder will learn to generate latent spaces that contain a(n approximately) discrete set of values. This is a great advantage, as it allows us to apply arithmetic coding on the latent space without it being passed through an operation with ill-defined gradients, such as quantisation is. Additionally, the framework can just as easily be used with an explicit quantisation built-in, where the latent space is trained against a discrete prior after being quantised. The difference between these two schemes is shown in
So far, all the entropy models and strategies we have described are fixed at inference time, that is, when we compress any image the entropy distribution will be the same.
An improvement over this fixed entropy approach is to incorporate some side-information in the bitstream: this side-information contains instructions on how to modify the entropy model for that particular image, so that a greater amount of flexibility is allowed, which results in a higher compression performance.
Traditionally, such side-information has been created with hyperprior networks in learnt compression. A hyperprior network predicts the moments of the prior distribution: for instance, it could predict the mean and variance of a normal distribution. This distribution is then used to entropy code and decode the latent space.
We provide a similar pipeline for framework 3, illustrated in
Additionally, we provide a different strategy. This strategy is based on the premise that for a fixed bitrate there are infinite probability distributions, and thus the objective of our model is to find the distribution that results in the highest reconstruction fidelity for a given bitrate.
This is achieved by setting the prior distribution to be categorical, i.e. a discrete distribution with a finite set as its support (e.g. the values {0, 1, 2, 3}) and arbitrary probabilities for each value (e.g. {0.1, 0.2, 0.5, 0.2}. These probability values can be either learnt over the training dataset, or predicted by a hyperprior model. This method is illustrated in
Note, in order for gradients to flow back from Wasserstein to the parameters of the hyperprior network (so that the hyperprior can learn to predict good probability values), it is required to backpropagate through the Sample operation, but sampling from a categorical distribution is normally not a differentiable operation with respect to the probability values.
We present a differentiable approximation of this operation. The Probability Mass Function (PMF) of a categorical distribution may look as in
First, we sample from a standard uniform distribution; secondly, we map each sampled value to categorical space with a piecewise linear function, where the width of each segment is dictated by the probability value of the categorical distribution. In order to discretise the values, we finally apply quantisation with a Straight-through Estimator to retain gradients for backpropagation. For an illustration of this process, see the example of
There is one more hurdle to overcome: since the probability values are predicted by the model, we need a transformation that maps R to the solution space of the under-determined system of equations below:
ΣiNpi=1
ΣiN−pi log2(pi)=B (9.13)
Where pi is the probability of each value in the distribution and B is a target bitrate that is known in advance and can be specified by the user.
We provide a transformation that contains an iterative method. The transformation algorithm is as below:
N
The final innovation we describe is an additional step on top of framework 3. The step comprises in taking the latent space of the auto-encoder and passing it to a normalising flow. The normalising flow is perfectly invertible, which means that if we take its output and pass it back to the flow, we will obtain the original input without any distortion.
We exploit this property by inserting the normalising flow between the latent and the entropy loss, i.e. we take the latent y, pass it to our normalising flow, obtain an alternative representation w and calculate our divergence measures on w instead of y. The pipeline is illustrated in
The normalising flow becomes part of the compression pipeline, so we only need to send w as bitstream and then reconstruct y from it at decoding time. The training process of such a system is described in Algorithm 9.6.
H×W×C
Using a normalising flow to further process the latent space allows y to retain spatial correlation information, while making w more similar to the prior distribution.
Below we enumerate the concepts we described.
First, we identify three image and video compression pipelines.
Furthermore, we identify two novel compression pipelines.
Finally, we identify two methods associated with concept 5.
Multi-task learning (MTL) with neural networks leverages commonalities in tasks to improve performance, but often suffers from task interference which reduces the benefits of transfer. To address this issue we introduce the routing network paradigm, a novel neural network and training algorithm. A routing network is a kind of self-organizing neural network comprising two components: a router and a set of one or more function blocks. A function block may be any neural network—for example a fully-connected or a convolutional layer. Given an input the router makes a routing decision, choosing a function block to apply and passing the output back to the router recursively, terminating when a fixed recursion depth is reached. In this way the routing network dynamically composes different function blocks for each input.
The introduction will cover the problem setting, the intuition behind the solution, and an overview of how it is implemented.
A general challenge of the neural networks is that they are computationally heavy and have a significant memory footprint. Therefore, they can not yet run in real-time on most consumer devices (edge devices). This execution-complexity remains a considerable challenge for the AI-based Compression pipeline, amplified by its strict real-time requirements of 33 ms per decoder pass. Note that we used the word “decoder” and not the “entire algorithm” in the last sentence. Compression can use asymmetric approaches in which encoding the data does not come with “too strict” time requirements, but decoding the data comes with harsh real-time restrictions. For instance, a 30 fps movie requires decoding times below 33 ms; a 60 fps movie requires decoding times below 16.3 ms.
Why are neural networks so slow to execute? There are two primary factors. First, neural networks require a tremendous amount of computations, often in the billions, for an inference pass. They have many (floating-point) operations (FLOPs) to execute; and are bottlenecked to the following fps:
Second, neural networks have a tremendous memory footprint and movement. This means that before data can be used for calculations, it has to be moved from one memory location to another. There is a limit to how much memory can be transferred per second, given by the memory-speed. Thus, a neural network fps is either constrained by its FLOP-limit or by its memory-limit (more often the case). The roofline model in
It is necessary to understand that the runtime issue will not disappear any time soon. While neural engines, specific chips designed for neural network execution-only, are becoming more abundant, they only increase the compute power. With more computing power, we can execute more (floating-point) operations (FLOPs) per second. However, the memory footprint issue remains, and there are not so many advances made on this topic recently.
It is well known in the Deep Learning community that we can build vastly different looking neural networks/architectures for a similar performance. However, all this achieves is to trade off memory footprint versus memory movement versus flops. For instance, for the same performance, we can build:
Neural Network Architecture Search (NAS) helps to find a good trade-off between these three properties. However, this does not help with runtime because we must reduce FLOPS, memory footprint and memory movement for a meaningful runtime reduction. So, what else can we do to get a runtime reduction? The answer is to use fewer operations of all types but use these operations more efficiently.
An example: Let's assume we train a neural network with a generative loss for compression. If we have a dataset that comprises faces and cars, we could train one generalised network for compression; or we could train two networks, one for the faces and one for cars. If we train two networks, we will see that our compression performance will be better than if we train only one network. Why is this the case? It happens because the available network operations specialise given their input data and become more efficient. Essentially, we re-formulated the problem of compressing “all images/videos” into the multi-class dataset problem of compressing numerous different classes of images/videos. If we go to the limit and train one network per class, we will get the maximum efficiency per operation in our network for its given class.
The specialisation described above does not help us reduce runtime due to the necessity of having N different neural networks for N data classes. If we have more neural networks, the memory footprint increases because of the network parameterisations/weights.
The real solution is to realise that even if we have multiple networks for multiple data classes, there is a high likelihood that “most” operations will be the same—only a few operations will actually specialise. Thus, instead of re-formulating the problem as a multi-class dataset problem, we can re-formulate the issue as a multi-task learning (MLT) problem. Therefore, our new tasks become to find one algorithm that can learn multiple tasks instead of multiple algorithms for multiple tasks.
This problem interpretation opens up the doors to use techniques from the MTL domain for AI-based Compression. Specifically, we are interested in using Routing Networks for our pipeline. Routing networks are neural networks that have multiple filter options per layer, and during inference, the data itself decides which path it wants to take.
A routing network [2] is a neural network for which the data flow through the network is not fixed but is routed. Let's assume we have a neural network that is composed of multiple layers. As a neural network is simply a chronological concatenation of numerous functions, we can write these layers as functions ƒ and the network as ƒ∘ƒ∘ . . . ∘ƒ∘ƒ. We call the first layer ƒX,1, the second layer ƒX,2 and the n'th layer ƒX,N. A normal neural network has a fixed function that is executed per layer. However, a routing network has multiple options we can pick from each layer. Let's assume there are M potential options to pick from in each layer. We call option m∈M in layer n∈N function ƒm,n.
An input to a routing network flows through all layers of the network {ƒx,
In a routing network, the router, or Routing Module, decides the path through the network. There are numerous options for designing a router, a global, a local or a semi-local one (see
While it might be obvious, note that routing networks come with an explosion of combinatorial possibilities for the final network. Let's assume we have an eight-layer network ƒ=ƒm,1∘ƒm,1∘ . . . ∘ƒm,1, and each layer has 16 possible choices; M={1, . . . , 16}. Then, the final network ƒ has 4 billion possible routes (MN=168=24×8 4, 294, 967, 296). Thus, routing networks' power comes from its combinatorical flexibility and from each route of the final network specialising in a certain input class.
For a more realistic illustration of how to use routing networks as the AI-based Compression pipeline, see
When training a routing network, there is one familiar and one new part. The familiar part is the network training through stochastic gradient descent: we get the gradients from a loss function and backpropagate them through all layers using the chain rule. Note that while the layers can change with each iteration, this training approach still works.
The more interesting question is, how do we train the routing module? The difficulty from training the router is that the ideal output is a discrete number m from our discrete set of options M. Thus, training the router is no longer in the domain of gradient descent training but requires reinforcement learning (RL) methods. Reinforcement learning faces the same problem of selecting an “action policy” from a discrete set of options given a reward function.
Finally, to build up some intuition about routing networks, the two training stages resemble two different trade-offs. Routing networks always have the exploitation versus exploration challenge. While the standard neural network resembles the exploitation part, the routing module choice resembles the exploration part. The challenge of training routing networks is to train both parts of the network in parallel.
Following is a short overview of the different types of RL approaches to training the routing module. In general, all possible RL methods can be used, with the most popular ones being:
1. Continuous Relaxation: Originally, we want the routing network to output a discrete choice. One approach is to relax this assumption during training and have the router output a vector of probabilities over the choices Pn. We can write the next layer as a combination of all possible choices with weight factors, as follows:
Using a continuous relaxation of the discrete space allows us to use gradient descent training methods for training the router. Note that in inference, we replace the continuous relaxation with a discrete choice using the maximum function. For instance:
Pn=Routern
Layern=Layermax(P
2. Discrete k-best choices: We can use all kinds of reinforcement learning approaches on either the best choice or the k-best choices of the router with k∈{1, . . . , M} (these are also called higher cardinality approaches). Amongst others, we can use: deep reinforcement learning, Markov decision process, Dynamic programming, Monte Carlo methods, Temporal Difference algorithms, model free RL, model-based RL, Q-learning, SARSA, SARSA, DQN, DDPG, A3C, TRPO, PPO, TD3, SAC.
Finally, it is essential to complement the training approach for optimising the loss/reward function for the routing module with a diversity loss. The diversity loss has to force the routing network to produce diverse outputs (e.g. routes). Routing networks tend to have much stronger exploitation than exploration, and if there is no additional loss forcing more exploration, they often collapse. In a collapsed routing network, the routing module only produces one route through the network; thus, losing out on its potential flexibility. For more information, see reference [1].
This is a short paragraph to link together some of our IP+concepts. Routing Networks are, in fact, a generalisation of Network Architecture Search (NAS). In a more broad sense, they generalise the concepts of NAS, Reinforcement Learning and One-shot Learning.
While NAS is about learning an efficient, fixed architecture for a neural network; routing networks are about learning multiple, efficient, flexible architectures for a neural network.
Thus, all the techniques used in NAS to make the network architecture selection more powerful, e.g. diversity in the layers, kernel size, operations and others, can also be used for routing networks. NAS is simple routing networks without routers: NAS+routing module+RL==Routing Networks. This is important as NAS and RL are their own gigantic domain of research. But we want to identify their methods for potentially training our routing network.
Now that we know the concept of Routing Networks, we must look at the Routing Module/Router. The job of the Router is to output data that will lead to a routing path decision. This could either be in the form of a discrete choice {1, . . . , M} or a probability vector over the choices {p(m==1), . . . , p(m==M)}. Afterwards, we need some strategy to pick a path/route given this output, as described in section 2.2.
There is no a priori restriction on what the Router must look like. For instance, it could be feature-based, neural network-based, or anything else, as would be clear to the skilled person.
Feature-based Routers are approaches that make a decision on classical computer vision features. For instance, we could use image statistics, histograms, gradient magnitude information, patch-wise features, feature-point extraction data such as FAST/SIFT/Others, Edge and/or Corners and/or Ridges detection, and many others. Feature-based approaches fix an image/video feature extraction method and then build a learning-based strategy (decision tree, regression, kernel PCA, support vector machine, or others) on top of this information. The benefits from these approaches are that the feature-extraction reduces the problem's dimensionality, from image space to feature space, and thus, leads to a massive Router runtime acceleration. Additionally, these approaches are usually resolution-independent and do not worry about different input data height and width.
Neural network-based Routers make a decision on deep-features and use the power of deep learning to make the Router a universal function approximator. The benefits are that it makes the Router more potent than classical feature-based strategies and that we are not using a priori information; everything is learned end-to-end. The drawback is that we require the output to have a fixed form, and for this case, standard neural network architectures are not resolution-independent. We address this issue in the next section.
Some more advanced insights: The encoder and decoder module in the AI-based Compression pipeline look like inverse functions of each other. E.g. it seems as if the decoder has to learn the inverse mapping of the encoder. Interestingly, this relationship appears to hold true for latent regions with high-probability mass, which means that the encoder and decoder's Jacobian are orthogonal at these points. If we use routing networks and make an encoder layer flexible, we need to pick the correct decoding layer to complement the encoding layer change to keep the orthogonality. Thus, there should be information flow between all routing modules, as the routing choice of a decoding layer should be influenced by the routing choice of the corresponding encoding layer. To facilitate optimal information sharing, local Routers have numerous skip connections between each other, as visualised in
There are multiple ways of making the Router resolution-independent while still using neural networks:
It is essential to give special attention to the router training to get the desired outcome. Routing networks, in general, have the challenge of balancing exploration versus exploitation. Specifically, it is a problem for the Router. If we train the Routing Network AI-based Compression pipeline naively, the Router will collapse and produce only one output for any input. Aka. the Router will 100% favour exploitation over exploration, and once it is stuck in a local minimum, it never recovers.
We can get around the collapse of the Router by facilitating exploration with a “diversity loss”. Basically, a diversity loss is an additional loss term that penalises the Router if it makes the same choice (selecting the same path) too many times. There are two choices for a diversity loss:
Before diving into the individual sections on lens distortion and intrinsic calibration, we will first address how a learning based approach is possible over flexible sets of input data. After all, we want to enable the user to select an arbitrary amount of images for the calibration process, thus giving him a trade-off between performance vs accuracy.
Given more than one input for a naive network h(·), the order of the inputs does matter for the outcome. For instance, h(x1, x2, . . . , xn)≠h(xn, xn−1, . . . , x1). This becomes a challenge when we try to apply deep learning to the problem of camera calibration for a collection of images. In essence, we want a network that considers the input images as a set, independent of their order. This property is called permutation invariance.
There exist a number of options one can use, including sorting the input, training an RNN with the desired property, or using a network structure that is symmetric with respect to the input. In fact, one may argue that using a symmetric network is theoretically the preferred option as it naturally ensures permutation invariance. It does not require additional augmentation and does not rely on any sorting heuristics. We can get a symmetric network hs(·) by sharing weights across all inputs that are being processed: hs(x1, x2, . . . , xn)=hs(xn, xn−1, . . . , x1).
However, such a architecture can still only process a fixed amount of input images. We need a network that is agnostic to the amount of input data, thus is invariant to the cardinality of the input data set. For this, we need to accumulate the output of the shared networks via a symmetric accumulation operator ⊕, for instance element-wise mean, element-wise maximum or summation. Afterwards, we can (optionally) process the accumulated data further via another network g(·), which now has a fixed input size. Thus, the core structure of our networks is as follows:
In this section well give an overview about the network architecture we use for both the lens distortion parameter estimation as well as the intrinsic parameter estimation. First we will give an introduction to the naive-version of the model, then introduce our contribution named global state fusion into it.
As our model needs to work with arbitrary-sized sets of input images, we first have shared networks which operate on different input images. We use the term shared to mean that all networks share the same weights across the same layers. These networks use blocks of densely connected convolutional layers, interleaved with downsampling layers (conv-layers, stride two) to reduce the dimensionality. Moreover, we use multiple skip connection via concatenations to aid proper gradient flow and reduce the vanishing gradient problem.
Second, after three downsampling operations, we fuse the outputs of the shared networks by averaging them as described in section “Cardinality Invariance”, followed by multiple fully connected layers to get our parameters.
We now extend our above-described model by introducing global state information into the shared networks. The key idea is that in the naive model the shared networks only have the information of one image, only fusing their individual knowledge at the end. In contrast, we want the networks to have global information at multiple points, and let them decide if they use it or discard it.
Therefore, after each block of conv-layers we average the output of all shared networks, and concatenate this average global feature map state to each one-image only feature map state. Thus, our shared networks do global state fusion multiple times during an iteration. Moreover, because we concatenate this information, the network can learn to which extent it wants to utilize it.
Let's name the output of conv-block j in the i'th shared network oij. Then before the next conv-block begins, we concatenate the network specific feature state oij with the global feature state õj, where by equals Σk=1n(okj). This operation keeps permutation and cardinality invariance. For an in-detail overview of the network see
In
Activation functions, or nonlinearities, are function mappings that transform the input space nonlinearly which are fundamental for effective information processing in neural networks. The Padé Activation Unit, or PAU, is a very efficient and generalised activation function that is capable of approximating many popular activation mappings currently employed in AI literature. This has still the capacity to be extended with variations in its parametrisation structure, evaluation algorithms and stability mechanisms. Furthermore, a multivariate PAU is also possible, extending the concept to process intervariable relationships between the input variables.
Neural networks are famously able to model highly complex relationships between observed data and latent features. It owes most of this to the activation functions, which grant the network its nonlinear modelling capacity. Activation functions, or nonlinearities as they are often called, are nonlinear function mappings that transform the inputted feature vector (or, simply, feature map) to an activated feature vector (or activation map). There exists a large variety of activation functions in deep learning literature, with names such as tanh, sigmoid, ReLU and Leaky ReLU being popular in the research field. Many of these differ in their functional expression, and there is little consensus of which activation function to choose for a given optimisation task. Moreover, if the activation function is not sufficiently flexible (or even fully static), this induces an arbitrary prior on the model. This can either aid the network in its task if the activation function is well-suited, or stifle its modelling capacity if it is poorly chosen.
A logical workaround would be to design and parametrise an activation function with ample degrees of freedom such that it can approximate most of the common activation functions to a sufficient degree, as well as embody less conventional or even completely novel nonlinear mappings. Ideally, the number of parameters for this should be small to facilitate modelling capacity and promote generalisation to the data. One elegant method for such a method is the Padé approximant.
The Padé approximant comprises a rational function
where pm (·) is a polynomial of order m, qn(·) is a polynomial of order n and x is some arbitrary input. In full, the Padé approximant can be expressed as:
As we shall see, with only a few parameters, the Padé approximant has the capacity to model virtually all activation functions that are used in neural networks within a reasonable range of operation. With such a generalised mapping, there is abundant design space for extending its parametrisation to encourage expressivity or to limit it to promote generality. This mathematical construct is the foundation of the provided activation function, the Padé Activation Unit or PAU, which we employ within the domain of AI-based data compression.
The focus here is to
Please see Section 2.2 for mathematical preliminaries.
Normally in neural networks, activation functions are interleaved with linear and convolutional layers, with the optional normalisation and pooling layer. This is the usual structure of a module such as the encoder or decoder (see
The innovation of the Padé Activation Unit will be clearly detailed in the subsections below, comprised by the following:
In a forward (propagation) pass, the data is processed sequentially through each neural network module and the forward functions of its singular components. In the case of the PAU as activation function, for the input hl, the forward functional expression is
where (m, n) is the order of the Padé approximant and in effect determines the parametrisation structure of ƒ(·), given by α={a0, a1, . . . , am} ∈m+1 and b={b1, b2, . . . , bn} ∈
n. Initially, it is assumed that a and b are global parameters for the activation layer (i.e. layer-wise activation function), but we shall see in later subsections that we can easily extend this to a channel-wise parametrisation structure.
This expression differs from the formal definition of a Padé approximant (Equation (11.2)) in that the terms in the denominator are kept positive with the absolute value operator. In its current form, the denominator of the PAU is guaranteed to be larger than one. This is to ensure the absence of poles causing numerical instabilities, which occurs when the denominator evaluates to (or approaches) zero.
Since the PAU consists of two polynomials, we can leverage efficient polynomial evaluation algorithms in our forward function. One such efficient algorithm is Horner's method, which expresses a polynomial as follows:
This algorithm requires m additions and m multiplications to run. Although it relies on serial execution where each addition/multiplication depends on the previous term, in most practical applications m is fairly low (see N) in memory (two polynomials plus the input).
N: input feature vector
m+1: PAU numerator coefficients
n: PAU denominalor coefficients
N: activated feature vector
1N is a N-dimensional vector of ones
Can be parallelised with line 9
Can be parallelised with line 6
Saved for backward pass
The backward function is defined to allow for the gradients to flow through the PAU to upstream modules, as well as update the parameters {a, b}, during network training. Automatic differentiation packages usually take care of the backpropagation process, however the backward functions can also be custom-defined, such that their computation can be optimised for (using CUDA kernels, for instance):
Here, hi,l is a scalar-element of the input vector,
These can also be evaluated using Horner's method or alternative polynomial evaluation strategies.
N: incoming loss gradient
N: outgoing loss gradient
m+1: loss gradients for PAU numerator coefficients
n: loss gradients for PAU denominator coefficients
The PAU can be parametrised such that its parameters are:
The partitioning can also be of finer structure, such as patch-wise or element-wise.
If the polynomial order of either pm (·) or qn(·) is large, we can employ Estrin's scheme and evaluate the polynomial in parallel (assuming that we have sufficient memory capacity). Given the polynomial or order m, we can rewrite it in a way that allows for parallelism
where r└m/2┘(x2) is a └m/2┘-degree polynomial in x2. Every bracketed term can be evaluated in parallel, hence the speed-up, and the scheme can operate further recursively, resulting in lower-degree polynomials in higher orders of x.
Alternatively, when the parametrisation is static and not under weight optimisation (so during deployment), Newton's method can be used to factorise the polynomials and simplify or approximate the functional expression of the PAU forward pass in order to optimise for the algorithmic evaluation speed and memory.
To avoid poles from arising in Equation (11.2), we implement “safe” PAU by restricting the terms in the denominator polynomial to nonnegative values. However, these can hurt expressivity and there may be better alternatives for the forward function that also safeguard against poles. Some of the stability mechanisms that are possible are:
Alternative absolute valuing: We can ensure that the denominator is always positive by taking the absolute value as such:
This is a more representative version of the Padé approximant formulation (Equation (11.2)) since it aggregates denominator terms before the absolute value is taken. However, the poles that otherwise would cause discontinuities are now manifesting as sharp peaks or troughs, which may disrupt the learning process.
Introducing b0 with positivity constraint: We can replace the one in the denominator polynomial with a bias term, b0, which we have to restrict to be larger than zero for stability purposes:
We can do this with a small constant stability term, E, to the absolute value of b0 such that no poles arise.
As of yet, the PAU has only been discussed in terms of the input directly, without modelling relationships between input variables. It would therefore be reasonable to consider the extension of the PAU to multivariate PAU, which consists of the quotient of two matrix polynomials
where the set of numerator coefficients, {α0, A1, A2, . . . , Am} are all matrices of dimensionality N×N except for α0, which is an N-dimensional vector. Likewise for the set of denominator coefficients, {B1, B2, . . . , Bn}, which are all
N×N. To keep dimensionality tractable, it is likely that this scheme will be employed for partitions of the input, such that N is for instance the number of channels. The matrix-vector product in each term, for example A2x2, can be expressed as a linear layer or a convolutional layer with weight matrix A2, for which the input elements will be taken to the corresponding power.
In fact, the multivariate PAU as formulated above generalises the concept of divisive normalisation, an operation in neuroscience, relating closely to how the visual cortex processes information. Assuming a bias term with a positivity constraint in the denominator, the multivariate PAU is very similar to the formulation of generalised divisive normalisation (GDN), a popular activation function in AI-based image compression
with the constraints that βi>0, ∀βi∈β and yi,j≥0, ∀γi,j∈Γ. If e is all ones, this formulation is easily encapsulated in the scheme of multivariate PAU.
In this section, we present the following concepts regarding the Padé Activation Unit as activation function. All concepts listed below are considered under the context of the wider domain of AI-based data compression.
A goal for the current state of the art neural image and video compression pipelines deployed for any type of streaming media is massively reduced latency and computational cost to manage the demands of modern and future VR streaming, cloud gaming, and any other innovative electronic media streaming service. Up until this point, there are no learned image and video compression pipelines capable of this feat. Here, we outline the building blocks for a neural image and video compression pipeline that runs wholly in the frequency domain to realize orders of magnitude lower latency and computational costs compared to any other published state of the art learned image and video compression pipeline.
Image compression pipelines powered by deep neural networks have in recent years been shown to consistently outperform the best traditional image compression codecs based on the High Efficiency Video Coding (HVEC) avd Versatile Video Codec (VVC). However for novel image and video applications such as live streaming, VR, AR and cloud gaming, satellite and medical imaging, 3D films, etc., these state of the art neural compression pipelines are still completely unsuitable due to strict latency requirements, high resolutions and slow run-time. To meet the stringent latency and compute restrictions of current and future media streaming services we present novel neural compression building blocks created to realize learned image and video compression pipelines in the spectral domain. A building block in state of the art neural image and video compression pipelines is the convolutional operation which constitutes close to all of the computational cost. Mathematically a convolution with kernel & with the original image ƒ(x, y) may be defined as in Equation (12.1).
Within the field of machine learning, and more recently deep learning, several efforts have been published improving the performance of convolutions that involve very large kernels using the mathematical theorem known as the convolution theorem, shown in Equation (12.2) where refers to the Fourier transformation. Briefly, the Fourier related transformations also referred to as integral transformations, are mathematical operations that are typically employed to shift to a more advantageous domain to operate within. The advantage for an image and video compression pipeline to transform into a spectral domain is evident in Equation (12.2)—pointwise multiplications in the spectral domain correspond to convolutions in the spatial domain, drastically reducing the number of floating-point operations.
{ƒ⊗g}=
{ƒ}*
{g} (12.2)
The traditional building blocks of the learned image and video compression pipeline such as convolutional layers, pooling layers, batch normalization layers, activation layers, etc., do not work within this domain. There is no neural image and video compression pipelines operating completely in the frequency domain. The domain is still unexplored—published research papers related to this area are surprisingly scarce. In this document, these types of neural compression pipelines of electronic media operating within the spectral domain may also be referred to as spectral compression pipelines. There are two large open questions within this niche field that most likely act as the bottleneck for academic and research interest:
Here we provide a novel toolkit of neural image and video compression building blocks to realize a neural image and video compression pipeline that runs completely within the spectral domain—the first spectral image and video compression pipeline.
The building blocks utilized in our spectral compression pipeline will be briefly discussed below.
We utilized a Fourier related integral transformation known as the Hartley Transformation. However, the specific integral transformation may not be important as long as the transforms are continuous (integral) transforms of continuous functions. Thus, the following methods may be applied in addition to the traditional Fourier Transformation: Hartley Transform, Wavelet Transform, Chirplet Transform, Sine and Cosine Transform, Mellin Transform, Hankel Transform, Laplace Transform, and others, for example.
As mentioned above, the traditional spatial neural activation functions that ensure non-linearity in typical learned compression networks may not be employed in a spectral compression pipeline. This is because in the spectral domain the effect of an activation function typically employed in the spatial domain is mathematically completely different. As such a variety of spectral specific activation functions were implemented, such as the spectral non-linear activation seen in Equation (12.3) below, where Fconv
Fact(x)=Fconv
An immediate limitation of spectral convolutions, based on Equation (12.2) is that pointwise multiplication between the kernel ω and image ƒ(x, y) necessarily requires that the shapes match. A method was implemented to ensure that the kernel ω and input ƒ(x, y) are of compatible shapes.
A spectral based learned compression pipeline may not necessarily require image scaling in the same sense as for traditional neural image compression. Nevertheless, as a way of achieving additional performance benefits a novel type of upsampling and downsampling was created specifically for the spectral domain, shown below in
Specifically in
Two varieties of receptive field based spectral image decomposition and image reconstruction for the spectral compression pipeline are discussed in this section.
The image decomposition is known as stacking: smaller image patches or blocks are stacked in a new dimension, whereas the image reconstruction is known as stitching. Specifically, a window of size WH, WW slides across the image based on a stride SH, SW. For each window position, a patch is created that is stacked in a batch dimension. The overlap between successive windows is based on the difference between the window size and the stride, for an example see
When reconstructing the image back together there are two methods of stitching. Firstly by stitching with the overlapping regions averaged (see
Neural Network Architecture Search (NAS) is an approach in which we attempt to remove human bias from decision-making regarding neural network architecture design. AI-based Compression is an approach in which we attempt to remove human bias from designing a compression pipeline to get to the next generation of compression technology. In their core approaches, AI-based Compression and NAS overlap. It is the next step to apply NAS to the network design of AI-based Compression to also remove human bias in the codec design.
Here we describe methods (NAS) of determining one or multiple candidate architectures for a neural network for performing AI-based Image/Video Compression for different use cases. These methods include: maintaining a sequence of neural layer (or operator) selection processes, repeatedly performing the candidate architecture forward pass, updating the Neural Architecture Search system by using the feedback of the current candidate sets, and selecting one, or a group, of candidates of neural architectures as the final AI-based Image/Video Compression sub-system; or as a particular function module for the final AI-based Image/Video compression sub-system.
The innovations include applying the NAS-process to
with the goals of getting
Let us define a few NAS related terms:
Here we apply NAS to optimal operator selection, optimal neural cell creation, optimal micro neural search, optimal macro neural search under the context of AI-based Image and Video Compression. We will consider different performance estimation methods and search space limitations to reduce search times; and use efficient search strategies.
For operator selection, the question is which function should we use at which position in the neural network. Given a fixed-architecture and a set of pre-selected operators, picking the best ones becomes a challenge. For example, suppose the set of possible operators is as follows:
O={convolution-layer-1×1, convolution-layer-3×3, convolution-layer-5×5, convolution-layer-7×7, activation-function-1, activation-function-2, activation-function-3, activation-function-4, Identity Function, Skip Connection, Attention-Module, adding bias, . . . }
Each time we select an operator in the network we must pick a specific function from O→ƒi∈O.
Once we have O defined, the question becomes, how can we train such a network, and how can we select one operator per function. In general, there exist two approaches:
Note that such a setup can give us additional possibilities to model non-standard loss objectives. For instance, we can associate auxiliary variables with the operators such as runtime, FLOPs, memory usage and others. Suppose we use these auxiliary terms in the loss equation. In that case, this gives us a straightforward way to optimise our pipeline for objectives such as runtime, computational complexity, memory usage, and others.
To search for an optimal Macro Architecture, we have two options: Either we start with a massive network and select strategies to prune connections/filters/weights (top-down approach) or build up an efficient architecture from scratch (bottom-up approach). There are also mixed-approaches which iterate between pruning and building (e.g. MorphNet).
We can combine any of these methods with AI-based Compression specific auxiliary losses. For instance, we select a pruning-approach and add runtime/memory/FLOPS/visual-quality/filesizes constraints to each operation and connection to train an optimal final model for our particular objective.
The Macro Architecture design's bottom-up approach relies on Supernetworks (Also called: controller networks, mother networks). We have an architecture we want to optimise (AI-based Compression pipeline), also called child-network, and a controller determining how good a child-network is. Known approaches are early-stopping criteria, building up result tables and using RL on these result-tables, using accuracy predictors. Examples include: FBNet, SparseNAS, and others.
AI-based compression uses neural networks that are trained to perform well and generalize across all inputs. However this leads room for improvement on a per-input basis (say, for one particular image or video). The role of finetuning is to improve an AI-based compression pipeline on a per-input basis. Here we outline several approaches: finetuning the latent variables; finetuning the decoder network's weights (parameters); and finetuning the decoder's execution path.
In a compression pipeline, an encoder sends a media file as a binary stream of bits. The encoder sends this bitstream to a decoder, which attempts to reconstruct the original media file from the bitstream. There are two competing tasks: on the one hand, the encoder wants to send as few bits as possible; yet on the other hand, the reconstructed media file should be as close as possible to the original file. This is the so-called “rate-distortion trade-off”: the compression pipeline must somehow minimize both rate (number of bits), and distortion, the reconstruction error between the original and decoded files.
Before delving into the rate-distortion trade-off, let's first outline a generic AI-based compression pipeline (see M(such as an image or a video file) in an M-dimensional space. In an AI-based compression pipeline, the encoder is a function E:
M
n, typically parameterized by a neural network. The encoder transforms the media into a latent vector y=E(x), y being a real-valued vector in an n-dimensional space. Usually n<M, so that the latent lives in a smaller dimensional space than the original media file.
Now, the pipeline must somehow turn the latent y into a binary bitstream. This is accomplished as follows. First, the latent y is quantized into a integer-valued vector ŷ. This quantized latent is given to a probability (entropy) model, which assigns a likelihood of each element in the latent occurring. These likelihoods are then sent to an arithmetic encoder, turning the likelihoods into a bitstream. The bitstream is what is actually sent by the encoder. On decode, an arithmetic decoder reverses the binarization procedure, taking binary values, likelihoods, and returning a faithful reproduction of ŷ. This recovered quantized latent is then sent through a decoder neural network, returning the final prediction=D(y).
How do we actually ensure that the prediction is as close as possible to the original input media x? Moreover, how do we control length of the bitstream (number of bits)?
These two issues are resolved during network training. Each of the encoder E, quantization function Q, probability model P, and decoder D may be parameterized by a large vector θ (sometimes called network weights). The parameters θ can be thought of as dials controlling the behaviour of the entire pipeline, and must be optimized. The parameters are chosen to minimize (1) the distortion between and x, and (2) the rate (length) of the bitstream.
During network training, the rate can be estimated without running the arithmetic encoder/decoder, using a rate estimation function R (see M×
M
is chosen to measure the difference between x and {circumflex over (x)}. These two quantities, rate aha distortion, are combined into a value called the loss,
:=R (ŷ)+λdist({circumflex over (x)}, x). A small loss means that the compression pipeline performs well. The coefficient λ>0 controls the relative importance of the distortion compared to the rate. Higher values of A mean that the loss cares more about minimizing distortion than minimizing rate; low values of A mean that the loss cares more about minimizing the rate (bitstream length).
The training procedure then attempts to minimize the rate-distortion loss over all input media files. That is, the training procedure tries to find a set of parameters θ that work equally well across all typical input media. In mathematical terms, the training procedure attempts to solve the optimization problem
The symbol is the expectation symbol, which means that the loss should be minimized in expectation (on average), over all possible media inputs. That is, during training we try to find parameters θ (for the encoder, decoder, quantization function, and probability model) that work well on average. Though obviously it is not possible to train a compression pipeline over all inputs (there are infinitely many), modern training methods are still able to find parameters θ that generalize well, so that the pipeline's compression performance generalizes to unseen typical images.
This points to the need for finetuning. The optimization procedure does not consider finding parameters that are particularly good for any one particular media file. Once training is finished, we have a set of parameters that work “pretty good”—the encoder, decoder, quantization function, and probability model all perform reasonably well on typical inputs. In other words, the compression pipeline generalizes well; but it is not specialized to perform superbly on any one particular image. It is a “jack of all trades, but master of none”.
The question that arises is then, can we somehow boost the performance of the compression algorithm on a case-by-case basis? I.e. given a particular input, can we design a per-image algorithm that improves the compression algorithm on a per image basis? This is the process offinetuning. Finetuning seeks to bring out extra compression performance (either in rate, or distortion, or both) on a per-input basis. Finetuning takes the already “pretty good” compression algorithm, which generalizes well, and somehow specializes the algorithm to perform very well on a particular input.
Example 1. To illustrate the idea, consider the following toy example, illustrated in
See for example
Here we are concerned with all possible ways of finetuning an AI-based compression algorithm. We consider three broad ways of finetuning:
In this section we discuss the technique of latent finetuning, and possible instances of this technique. The basic framework algorithm for latent finetuning works as follows (refer to Algorithm 14.1). For a particular input x, the finetuning algorithm begins by initializing with the quantized latents first produced by the encoder, setting ŷ0=Q(E(x)). The initial latents ŷ0 are generic latents, produced by the compression algorithm optimized to perform well on all possible inputs. These generic latents will be modified (finetuned) in some way to improve the compression performance. In a loop, the latent finetuning algorithm iteratively improves the latents, progressively perturbing the latents so that some performance metric
M, encoder ∈ :
M
n,decoder D :
n
M,
:
M ×
M ×
M
(x, ŷk, {circumflex over (x)}k)
of the compression pipeline improves. The performance of the compression pipeline is measured by a finetuning loss , which could for example measure:
At each iteration of the loop, a perturbation is generated, which is used to modify the latent. Perturbations are generated to improve the finetuning loss in some way. The prediction {circumflex over (x)}k is created from the current latent (which may be needed to determine how well the new latent performs, e.g. with respect to distortion). The iteration the begins anew. The loop ends when the latent is deemed optimal in some sense, and returns the finetuned latent.
Why is latent finetuning necessary? Remember that in a trained AI-based compression pipeline, the encoder E is optimized to perform well on all typical inputs; E is generalized, not specialized to the particular input at hand. Thus it is very likely that the initial latent ŷ0=Q(E(x)) is not the best latent for the particular input x, and that we can improve on the latent in some way. Notably, changing the latent ŷ may come with no increase to the bitstream length: no additional information is needed if we perturb ŷ in a sensible fashion (compare this with the methods of Sections 14.3 and 14.4, where extra information [bits] must be sent).
In mathematical language, the finetuning algorithm detailed in Algorithm 14.1 seeks to solve the following optimization problem
where is the finetuning loss; and {circumflex over (x)}=D(ŷ) is the output of the decoder.
The latent finetuning framework can be fleshed out in various ways. For example,
The remainder of this section will flesh out these various modifications to the latent finetuning framework.
The variable ŷ is ultimately the variable sent (via the probability model, c.f.
This is sometimes called STE finetuning (STE being short for “straight through estimator”). However, note that ŷ is the output of a series of transformations, in particular ŷ depends on y and x. Therefore, rather than finetuning ŷ itself, we could “pull-back” the optimization to any one of the prior variables in the compression pipeline (e.g. any one of the prior variables to ŷ in
So for example, rather than optimizing ŷ, we may optimize y. The mathematical problem to be solved then is
Note the subtle difference to Equation (14.3). The optimization variable ŷ has been replaced with y. And in the finetuning loss, we have made the relationship between ŷ and y clear by explicitly setting ŷ=Q(y). How does this change affect Algorithm 14.1? Because now the optimization is performed on the unquantized latent, initialization begins by setting y0=E(x). Perturbations will be generated for the variable yk, and the update will be yk+1←yk+p. Wherever ŷk is needed in the algorithm, it will be calculated on the fly as ŷk=Q(yk).
As another example, the variable to be optimized could be the input to the entire compression pipeline. Let's denote a generic input as {tilde over (x)}, and the specific image at hand as simply x. The mathematical problem to be solved then is
The optimization variable here is {tilde over (x)}, which effectively parameterizes ŷ via a pull-back ŷ=Q(E({tilde over (y)})). The changes to the framework Algorithm 14.1 are that: (1) initialization begins with {tilde over (x)}0=x; (2) perturbations are generated for {tilde over (x)}, so that the update rule is {tilde over (x)}k←{tilde over (x)}k+p. Whenever ŷ is needed, it is calculated as ŷk=Q(E({tilde over (x)}k)).
The finetuning loss, which measures how well the latent performs, plays a critical role in the finetuning algorithm. The finetuning loss may be used to generate the perturbations of the latent in the latent. In addition, the finetuning loss may be used to decide when to stop the iterations of the finetuning algorithm. The finetuning loss could measure
There are many possibilities for what the distortion metric in the finetuning loss. Possibilities include
Algorithm 14.1 provides a framework for perturbing the initial latent vector ŷ, however it lacks details of how the perturbation is actually constructed. There are many possibilities; this section will discuss some possible strategies for perturbing the latent.
The perturbation vector p of Algorithm 14.1 may be found by using a 1st-order optimization method, which solves the particular minimization problem (e.g. equations (14.3), (14.4), and (14.5)). A 1st-order optimization method is any method that approximates the loss (in this case, the finetuning loss), using the loss value at a point, and its gradient at this point (the direction of steepest ascent). So for example, the gradient descent method could be used to update the latents:ŷk+1=ŷk−τ∇ŷ
(x,ŷk,{circumflex over (x)}k) (14.6)
Here ∇ŷis the gradient of the finetuning loss, with respect to the latent variable ŷ. To be explicit, the perturbation is given by p=−τ
(x, ŷ, {circumflex over (x)}k). The scalar τ is a small parameter that controls the magnitude of the perturbation, the so-called “step-size”. τ can be calculated using any step-size rule.
This is just one of many 1st-order optimization methods. Other examples of 1st-order optimization methods that may be used are: Adam; any accelerated 1st-order method such as Nesterov's momentum; and proximal gradient methods.
The 1st-order optimization method can be applied to any one of the variables discussed above in the latent finetuning optimization methods (e.g. problems (14.3), (14.4), and (14.5)).
2nd-order optimization methods may also be used. A 2nd-order optimization method is like a 1st-order optimization method (using the loss value and its gradient at a point), but also uses the Hessian (the matrix of second-order derivatives of the loss). In a 2nd-order optimization method, the perturbation p is chosen to minimize a 2nd-order approximation of the finetuning loss
Here ∇ŷ2is the Hessian of the finetuning loss. The perturbation p is chosen to be no larger than some step-size threshold τ (the search radius).
The expression p can be evaluated using efficient automatic differentiation techniques such as the Hessian-vector product.
Note that the perturbation may also be constrained so that the update to the quantized latents is still an integer-valued vector. In this case, the problem is a quadratic-integer valued problem, which can be solved using algorithms for the Closest Vector Problem.
The latent perturbation need not be generated explicitly from local approximations of the finetuning loss (as in the previous two subsections, which used gradient and Hessian information). The perturbation could be chosen as a vector from a random distribution. This is the idea behind Monte-Carlo methods and their many variants.
M, encoder E :
M
n, decoder D :
n
M,
:
M ×
M ×
M
(x, ŷ′, {circumflex over (x)}′ )
(x, ŷ′, {circumflex over (x)}′ ) satisfies improvement criteria then
The general procedure is outlined in Algorithm 14.2. At each iteration, the perturbation is sampled from a probability distribution P, defined over the space of integer-valued vectors.
A new candidate latent ŷ′=ŷ+p is set. Then, this candidate is checked to see if it improves the latent finetuning loss in some way. If it does, then the candidate latent is accepted as the new latent. The loop begins anew, until a stopping criteria is reached.
There are several variants to this algorithm:
The latent perturbation can be chosen to only affect a small portion of the latent vector. At the extreme end, the perturbation could be applied to only one pixel (element) in the latent vector. In this scenario, it may not be necessary to run the entire perturbed latent ŷ k through the decoder network (to check the prediction {circumflex over (x)}k's quality). Instead, only a small portion of the latent may be needed: all those pixels adjacent to the perturbed pixel, in the receptive field of the perturbed pixel. The receptive field of the perturbed pixel are all latent pixels needed to compute prediction pixels that are influenced by the perturbed latent pixel.
When only a small portion of latents are needed each iteration, the entire finetuning process can be parallelized. That is, on each iteration a “batch” of many small subsets of the latent vector are processed in parallel. For example, in Algorithm 14.2, at each iteration, a batch of single pixel perturbations could be generated in parallel. Each of these perturbations may then be tested to see if they improve the finetuning loss (where only the local receptive field is checked, for every single-pixel perturbation in the batch). Only those single-pixel perturbations that improve the loss are accepted, and are used to update the latent.
The latent perturbations may be modeled as a Gaussian process. In this scenario, the perturbation itself is modeled as a parameter, to be learned as a Gaussian process. The perturbation is assumed to follow a multivariate Normal distribution. The Gaussian process modelling the perturbation is learned by updating the kernel function of the Gaussian process.
This is similar to interpreting the perturbation as hyperparameters from a given set, and learning these hyperparameters with a Gaussian Process. This can be viewed as an image-specific, natural extension of learning other hyperparameters, e.g. the learning-rate and/or the weight-decay, with Gaussian Processes. The details of how to execute this “smart” hyperparameter search using GP is common industry knowledge. Note, that for scalability we need overlapping GPs, Mixture-of-Experts (MoE) GPs or other modern techniques to make the computations feasible in practice.
In a compression pipeline, latent values that are zero are extremely easy to compress, and come with almost no bit cost. Therefore, it may be desirable to encourage the latent vector to be as sparse as possible (A vector is sparse when it is made mostly of entries with value zero).
Thus, sparsity inducing methods may be used on the latent vector. For example, the following optimization problem may be solved
where ∥ŷ∥0 is the l0 counting “norm”, which counts the number of non-zero elements in ŷ.
Several optimization strategies can be used to tackle this problem. For instance, hard thresholding may be used;
This function zeros any values that have magnitude less than s, but leaves all others untouched. Then an example of a hard-thresholding update rule is to set ŷk+1=(ŷk+E(x)−E(D(ŷk))). Effectively, this update rule pushes the latents towards sparsity while still keeping distortion of the prediction small.
Another strategy is to relax the counting norm to the 1 norm, ∥y∥1=Σ|yi|, so that the sparsity inducing optimization problem is
A method of tackling this problem is via iterative shrinkage. Define the shrinkage operator
An iterative shrinkage update rule would set ŷk+1=(ŷk+E(x)−E(D(ŷk))). This too has the effect of sparsifying the latent space, while still maintaining minimal distortion.
The problem of latent finetuning can also be cast as one of Reinforcement Learning. In this setting, the construction of the latent perturbation is tasked to an agent, which for example could be another neural network. In a Reinforcement Learning setting, the agent takes an action, which in our setting is the choice of perturbation. If the perturbation chosen by the agent improves the finetuning loss, the agent receives a reward. If, on the other hand, the agent worsens the finetuning loss, the agent receives a penalty. The agent's goal is to maximize its rewards (and minimize its penalties). A Reinforcement Learning algorithm is used to train the agent to make good actions (good latent perturbations).
Once the agent has been trained, it can be deployed into an AI-based compression pipeline to finetune the latent variable. So for example in Algorithm 14.1, the agent will be responsible for updating the latent ŷk with a choice of perturbation p. Note that the reinforcement learning algorithm could also be used to update any of the “pull-back” variables, such as y or x, parameterizing ŷ.
Latent finetuning shares many similarities with the Deep Learning subfield of adversarial attacks. Research has shown that neural networks can be extremely sensitive to tiny perturbations to their input (for example, an input image; or in our case, the latent vector). In the subfield of adversarial attacks, perturbations are created to break the network in some way. For example, if the network's job is to classify an image (say, as a cat or a dog), then an adversarial attack could be a tiny perturbation, imperceptible to the human eye, that causes the network to mis-classify the input image. It turns out that creating these types of adversarial perturbations is often surprisingly easy.
Most often, the route to creating an adversarial perturbation is (as is common in machine learning) through a loss function. The loss function measures the performance of the neural network (smaller loss values meaning that the network is performing well). In adversarial attack—unlike in latent finetuning—the perturbation must make the performance of the network worse. Therefore, perturbations are created which maximize the loss. Typically, there will also be a constraint keeping the perturbation imperceptible to the human eye.
Thus, there are many similarities between adversarial attacks and latent finetuning. Whereas an adversarial attack seeks to maximize a loss, latent finetuning seeks to minimize a performance loss. Both however attempt to keep perturbations minimal in some way, so that the perturbations effect is not (or barely) visible to the human eye.
Therefore, any adversarial attack method can be used for latent finetuning, simply by using a finetuning loss that should be minimized (rather than maximized). In a certain sense, latent finetuning is a kind of “reverse adversarial attack”, or a “friendly attack”.
Examples of adversarial attacks that can be used for latent finetuning include
The behaviour of the decoder D, which takes the latent variable ŷ and outputs a prediction {circumflex over (x)}, is controlled by the parameters of the decoder's neural network. These parameters include:
The parameter α could act on a particular channel; could be a single scalar; or could act on a per-element basis.
The parameter δ controls the bin size of the rounding function. The parameter δ could act on a particular channel of y; could be a single scalar; or could act on a per-element basis.
After a compression pipeline has been trained, in a standard pipeline all of the parameters (denoted θ) of the decoder are fixed and immutable. The innovation of functional finetuning is that in fact, some or all of the parameters of the decoder may be modified on a per-input basis. That is, a functional finetuning unit (see
Of course, since the additional parameters ϕ are calculated on a per-input basis, and so they must be encoded in the bitstream in some way, as meta-information in the bitstream. Thus, the additional parameters ϕ come with the cost of additional bits. However it is hoped that the extra information needed to represent ϕ is compensated by improvements to the bitstream length of ŷ and/or with a reduction in distortion between x and {circumflex over (x)}.
The additional parameters ϕ may be encoded in the bitstream in one of several ways.
Other ways to use the additional parameter in the decoder include:
An illustration of how ϕ could be used is the following. Suppose ϕ could be drawn from a finite set, so that it can be encoded using a lossless encoder. Then, for a given ŷ (the quantized latent produced from x by the encoder), ϕ could be chosen to minimize the rate-distortion trade-off (where now rate measures the additional bitstream length of encoding ϕ):
Here R(ϕ) is the rate (bitsteam length) of ϕ, and {circumflex over (x)}=D(ŷ; θ, ϕ) is the output of the decoder. In this example the decoder is parameterized by both the general parameters θ (fixed after the compression pipeline has been trained), and ϕ (which are chosen on a per-input basis according to the optimization procedure).
Note that finetuning the decoder D (this section) and finetuning the latents (Section 14.2), are not mutually exclusive procedures, and can complement each other.
A convolutional neural network is made up of a series of convolutional operations, and activation functions. Let's let the input to one of these convolutional operations be a tensor of shape Cin×H×W. Given an input x and a convolutional kernel K with Cin input channels, and Cout output channels, the convolutional operation can be written as
That is, the j-th output channel is the sum of convolutions over the input channels. This can be viewed as a fully-connected network over the input channels: the output of each layer depends on all previous channels. See for example
The idea of this section is to sparsify the convolutional kernels of each layer, on a per-input basis (to sparsify means to render an object sparse, in that it has few non-zero elements). This means that, given a fixed input to the neural network, many of the channel weights will be inactivated, and not used in the computation. This can be done for example with a binary mask M, where M has shape Cout×Cin. I.e. mij∈{0, 1}. Then,
If the mask has many zero elements, this can massively reduce the number of computations needed in each layer, for only channels with non-zero masks will be used in the computation. This is illustrated for example in
Importantly, the mask can be optimized on a per-input basis. For example, the mask can be chosen to improve the rate-distortion loss of the input. The optimization of the mask can be done in several ways:
The binary mask must be transmitted in the bitstream. The binary mask can be encoded with any lossless encoder.
Note that choosing the optimal mask is itself a non-linear operation. Therefore, it may be possible to use a decoder D without any other non-linear activation functions. Once the mask has been chosen, the masked decoder network is a series of linear transformations, which may massively speed up decode time.
The current media compression advances of state-of-the-art AI-based image and video compression pipelines are still severely limited by the computational demand of these algorithms. Practical use of better compression methods requires these approaches to run in real-time, defined as at least 25 frames per second decoding time. Up until this point, there are no learned image and video compression pipelines capable of this feat. In fact, current AI-based compression approaches are at least 1,000×too slow and computational-heavy to run in real-time [3]. Here novel methods of offloading decoding cost to the encoding phase for a learned image and video compression pipeline are presented. Our innovation uses metadata to transform the conditioned decoder into a linear function to realise real-time decoding times for high-resolution data. These methods may be collectively referred to as KNet.
Lossless data compression is about minimising the amount of information required to explain the data. The data could be an image, video, VR-data, AR-data, satellite data, medical data, text et cetera, so long as it can be represented in some latent, compressed form, that holds the same amount of information as the original data. Lossy compression is the same as lossless compression without the requirement to recreate the original data perfectly but allowed to have some distortion in the output. Our described innovation can be applied to lossy and lossless compression.
Compression algorithms have an encoding part which compresses the data, and a decoding part which decompresses the compressed data into the original data (with some distortion). Compression codecs are well-researched and standardised compression algorithms.
We call all compression codecs that do not utilise neural networks “traditional compression” approaches. The vast majority of all codecs, and all commercially available codecs, are from the traditional compression approach. In the past three years, there is a new class of compression algorithms being researched. These new algorithms are based around neural networks, have entirely different properties compared to the traditional approaches, and we call them “AI-based compression” methods.
Recently, AI-based image and video compression has shown tremendous promise and is already at a maturity level to outperform traditional-based image and video compression methods such as JPEG, JPEG2000, WEBP, BPG, HEIC, HEIF, H.264, H.265, AV1, H.266 [4].
A remaining challenge to transition this technology from “research” into an application is the issue of runtime. An image and video compression codec that cannot run in real-time is not a viable product. Especially noteworthy is the decoding time; users expect to see content, e.g. movies, with 25/30/60/90 frames-per-second (fps). Thus, the decoding time of a compression algorithm must be under 40/33.3/16.6/11.1 milliseconds per frame, respectively, to satisfy the demand of the user.
Note that the algorithm's runtime and the decoding runtime are related but not the same properties. The codec's overall runtime is the encoding time plus the decoding time. The encoding time is measured as the time it takes to compress raw content into the compressed bitstream. The decoding time is measured as the time it takes to decompress the bitstream into the final output content.
In the vast majority of the use cases for image and video data, decoding time is significantly more important than encoding time. This asymmetry is reflected by the asymmetric encoding-decoding process of traditional image and video compression approaches. Algorithms such as WebP, HEIC, HEIF, HEVC, AV1 and others, have 100×-1000×runtime differences between encoding and decoding, with decoding being quick and encoding being slow. For instance, for the use case of video-on-demand, Netflix states [5] that a 100× complexity increase in encoding would be acceptable without causing any problems, given that it is accompanied by better compression performance with adequate transmission time of the data to all end users.
The current state-of-the-art neural networks used in AI-based compression approaches do not utilise this asymmetry property and are mostly symmetric. Thus, the state-of-the-art of AI-based compression approaches have similar decoding and encoding times, both of which are, in most cases, too slow to be marketable.
Every compression codec faces the challenge of balancing runtime and performance this is especially true for AI-based compression. AI-based compression builds its framework around the usage of neural networks, and neural networks require immense computational efforts. To give three examples:
First, using the performance-optimised AI-based compression pipeline from [6], the runtime is 230 ms for a 768×512 image on a non-mobile CPU, with smartphone CPUs being 5×-10× slower than non-mobile CPUs. Extrapolating this data to higher-resolutions, we can find approximate decoding time for various resolutions in the table below:
Thus, the “efficient” AI-based compression pipeline is 150×(non-mobile) to 750× (mobile) times too slow to be used in practice for 4K 30 fps video. And the “efficient” AI-based compression pipeline is 600× (non-mobile) to 3,000× (mobile) times too slow to be used in practice for 8K 30 fps video.
Second, we can calculate the number of floating-point operations (FLOPs) required by the decoding neural network using the architecture described in [4]. The decoding neural network requires 48 TFLOPs. Modern smartphones can, at best, process up to 100 GFLOPS to 1 TFLOPs. Thus, running a 4K decoding pass with 30 fps would require 1,440 TFLOPS or 1,440× the processing power of modern smartphones, assuming 100% of the theoretical FLOP-capacity can be used.
Finally, we can look at the decoding times of different AI-based image compression approaches of the CLIC challenge [7]. The leaderboard of the CLIC challenge 2020 shows the average decoding times of different compression approaches over the CLIC Validation image dataset, consisting of 102 mobile and professional photos of varying resolutions, ranging from 384 to 2048 pixels per dimension. While BPG, a traditional compression approach, requires 696 ms, the AI-based approaches require on average roughly 100s (some algorithms up to 300s) per image. Thus, the AI-based methods are 150×-450× times slower than the traditional approaches. This comparison would be even worse for practical use cases, as the BPG algorithm was executed on a CPU, whilst the AI-based algorithms were executed on computationally-powerful GPU platforms. In practice, GPUs are rarely available, and CPUs are up to 10×-100× slower than GPUs for neural network executions.
In short, current AI-based compression pipelines can not be run in real-time. In fact, decoding times are multiple orders of magnitude too slow for even 30 fps-streaming. We need radical change to make it work.
A linear function is a function ƒ(x) for which Properties 15.1 and 15.2 below hold:
ƒ(a+b)=ƒ(a)+ƒ(b)(15.1)
ƒ(λ·a)=λ·ƒ(a)(15.2)
We can represent any linear function as a matrix multiplication and an addition. For an input x ∈N×1, a weight matrix W ∈
M×N, a bias b ∈
M×1, and [·] being the standard matrix-vector multiplication operator, a generalised formulation of linear functions is thus:
ƒ(x)=W·x+b(15.3)
A striking property of linear functions is that the function-wise composition of two, or multiple, linear functions remains a linear function. For instance:
ƒ(x) is linear,g(x) is linear→h(x)=g(ƒ(x))=(g∘ƒ)(x) is linear (15.4)
Mathematically, with the above-mentioned matrix-bias-vector notation, this is easy to prove. Let x be the input, Wƒ and bƒ be the parameters for the first generalised linear function ƒ(·), and Wg and bg be the parameters for the second generalised linear function g(·). Then, the function composition (g∘ƒ)(·)=h(·) can be written as
where the function-wise composition of the two linear functions ƒ and g give rise to a new linear function h with parameters Wh and bh.
Nonlinear functions are all functions for which either of the Properties 15.1 or 15.2 do not hold. Nonlinear functions have significantly higher expressive power and modelling capabilities than linear functions. For a conceptual intuition, linear functions are only able to represent straight lines, whereas nonlinear functions can also represent curves, something which is much more difficult for linear functions. Therefore, a nonlinear function has much more modelling flexibility than a linear one.
For example, in
A neural network is conventionally comprised of alternating linear and nonlinear operations cascaded iteratively. Most neural networks are based around the repeating structure:
As illustrated for example in
Please note that there are numerous ways of expressing n-dimensional convolution operations. We can either use the convolution symbol [], or instead, flatten the input and use the matrix-vector product. Both expression are equivalent; after all, a convolution is a linear function, and thus, can be written in the generalised linear function format mentioned earlier.
In neural network semantics, a nonlinearity is interchangeable with the term activation function, which inherits its name based on the idea of the action potential firing within biological neurons. Typical nonlinear activation functions include sigmoid, Tanh, ReLU and PreLU. Training a neural network can be seen as fitting a nonlinear function to input data.
It is essential to understand the significance behind nonlinearities inside a neural network. The nonlinear operations that follow the convolution and bias operations are the reason for the nonlinearity of the entire network and are the reason for the expressive power of neural networks. It is commonly accepted that the more nonlinear a neural network is, the better its problem modelling capacity. Even further, the entire proof of neural networks being universal function approximators relies on the nonlinearity, and the proof does not work without these operations [8][9].
In short, nonlinearities are an essential part of neural networks, and cannot be removed without significant penalties to the network's expressive power.
Mathematically, we can write a neural network with N repeated convolution-bias-activation structures as:
ƒN(WN·ƒN−1(WN−1·( . . . ƒ1(W1·×+b1))+bN−1)+bN) (15.7)
With ƒi(·) being the ith layer nonlinearity, Wi representing the ith layer convolution and bi representing the ith layer bias.
Let us now assume that we would remove the nonlinear operations from a neural network. Then the typical neural network chain would devolve to a sequence of purely linear operations:
Mathematically, we end up with a composition of linear functions. Owing to (15.5), we can thus write the entire network as one single linear function:
Proof:
ƒ(x)=ƒN(WN·(ƒN−1(WN−1 . . . (ƒ1(W1·x)+b1))+bN−1)+bN)
⬇
Remove nonlinear functions ƒN(·)
⬇
=WN·(WN−1·x+bN−1)+bN
=(WN·WN−1)·x+(WN·bN−1+bN)
=Wnew·x+bnew (15.8)
Thus, a purely linear N-layer neural network is equivalent to a 1-layer neural network. Mathematically, such a network is equivalent to using multivariate linear regression. Since this neural network has demoted to a purely linear function, it loses expressive power. However, thanks to the ability to squash a chain of linear functions into one linear composition function, the number of operations necessary to perform a forward pass has been reduced dramatically. As a result, the network can gain significantly in runtime performance, since a linear single-layer network can be executed much faster, and with substantially less memory footprint (memory access time) than an N-layer network. In essence, choosing the network complexity induces an implicit trade-off between predictive performance and runtime. We visualise this trade-off in
Nonlinear Neural Network: conv→bias→nonlinearity→conv→bias→non-linearity→ . . .
Linear Neural Network: conv→bias→conv→bias→conv→bias→ . . .
The current state-of-the-art in neural network architecture design is to use element-wise nonlinearities. In other words, every element in an input tensor is activated independently from one and another, and only depends on its current value as it is passed into the activation function.
Instead of thinking of an element-wise nonlinearity as a function, alternatively, we can think of it as an element-wise multiplication with a tensor that is dependent on its input. For instance, without loss of generality, the ReLU function in Equation (15.9) can be thought of as an element-wise multiplication between the input x and a mask R, consisting of is and 0s, that has been conditioned on the input x (15.10). Thus, ReLU can be restated as (15.11), where ⊙ is the element-wise multiplication operation.
With this interpretation of activation functions, our innovation is to replace the element-wise nonlinearity with a convolution operation whose parameters have been conditioned on its inputs. The values attained by the parameters of these convolutions, comprised by a convolution kernel, are dependent on the input with the dependency being fully described by a nonlinear function.
Let's assume we have a neural network with two convolutional layers represented by W1 and W2. We will ignore the bias without loss of generality. The exact definitions of the kernel weights of W1 and W2 determine whether the neural network is a linear function or a nonlinear function. If W1 and W2 are both operations with fixed convolution kernel, e.g. the kernel weights are constant across all input, the network is linear. However, if one of the operations, let's say W2(·) without loss of generality, is dependent on the input, the situation changes. If the function determining the weights of W2, namely W2(·), is nonlinear, then the neural network is nonlinear. If not, then the network is linear.
W2·W1·x is linear⇔If W2and W1are constant
W2(W1·x)·W1·x is linear⇔Only if W2(W1·x) is linear
W2(W1·x)·W1·x is non linear⇔Only if W2(W1·x) is non linear (15.12)
Chaining multiple layers of a neural network together with the novel convolution nonlinearity, and ignoring the bias for simplicity and without loss of generality, we get:
This chaining procedure can be termed kernel composition, since the resulting kernel from the sequential convolution kernels is a composite convolution kernel encapsulating all of its constituent kernels. The algorithm of this procedure can be seen in Section 15.3.4 and an example visualisation can be seen under Section 15.5 in
The mathematical structure of the proposed nonlinearities can be expressed as a linear function (such as a convolution operation), whilst the values they attain have originated as a result of nonlinear transformations on the input.
With the above-described innovation, the nonlinear convolution operation, what do we win? It is crucial to note that the entire neural network, composed of convolutions, nonlinear convolutions and biases remains a nonlinear function, with high predictive power but slow runtime.
However, if we condition the neural network on the convolution-kernels of the nonlinear convolution, we end up with a linear network, with the power of a nonlinear network.
The entire network(encoder and decoder)network⇔is a nonlinear function
The encoder network⇔is a nonlinear function
The decoder network⇔is a nonlinear function
The decoder network conditioned on meta-information⇔is a linear function(15.14)
Mathematically, this is easy to see, as the conditioning simply crosses out the input-dependencies:
The innovation is to use nonlinear convolution in the Decoder of the AI-based compression pipeline. During the encoding path, the user predicts the nonlinear convolution kernels.
Additional to the compressed bitstream, the encoding-user sends these kernels as meta-information to the receiving user. The receiving user uses the additional meta information and conditions the decoding network on that information, resulting in, from his point-of-view, a purely linear neural network.
Thus, we can combine the predictive power of nonlinear neural networks, with the runtime benefits of purely linear neural network—All at the cost of some additional meta-information.
We use nonlinear convolutions as an operation that is nonlinear if it is unconditioned, but which becomes linear once it is conditioned on appropriate metainformation.
We use the nonlinear convolution as an example for numerous potential classes of operations with this property, as it showed the best performance in our tests. However, the innovation comprises all classes of operations with these properties and not merely nonlinear convolutions. For instance, the innovation of conditioned linear decoders will also hold true if we replace the nonlinear convolutions with nonlinear element-wise matrix multiplication; or nonlinear matrix multiplication; or a nonlinear addition operation. The innovation is about the conditioning the make a nonlinear function linear in the context of neural networks; not about the exact way we use doing it.
Let's assume we have a function spacewhich we can describe as the union of two disjoint sub-spaces
L and
NL·
L being the set of linear functions in
,
NL being the set of nonlinear functions in
.
=
L
NL
L
NL=θ (15.15)
Functions in TL have fast execution time but limited expressiveness, whereas functions in NL have slow execution time but strong expressiveness. Our innovation proposes an efficient way of finding a function ƒ in
which is in the set
NL, but which is part of the set
NL when conditioned on additional meta information m.
ƒ∈and ƒ|m∈
(15.16)
Table 15.1 and Table 15.2 show an example layout of the network architectures used during training and inference of KNet.
B×C×H×W
C×1×kH×kW
+
C
C
C
3×C
3
One of the problems with the KNet-based architecture is that it is incredibly difficult to train in an end-to-end fashion. This challenge originates from the KNet module requiring a stable input-distribution to train, but the input to the KNet module is constantly changing via backpropagation in an end-to-end setting. This section provides details on how we can train the KNet module in a non-end-to-end fashion.
There are two ways of doing so:
For example, a linear regression analysis produces the optimal filter that the KNet module ideally would learn, and using this optimum as an initial proxy for our actual KNet module prediction aids the subsequent training process of actually optimising the KNet module with a frozen autoencoder backbone.
The challenge with the second point described immediately above is that linear regression only works with the assumption of no multicollinearity and assuming it processes semi-sensible inputs (ensuring stable training throughout). Generally, we cannot guarantee either of these. However, there are ways that can help us in the process. For example, for training stability, we can start o training with both a conv-gen and a conv-reg simultaneously, operating in parallel on the same inputs and yielding two different outputs and therefore two different loss components, gen and
reg, respectively. The final loss metric
is hence a weighted sum of these two as such:
=α
gen+(1−α)
reg (15.17)
Initially, the weighting factor α∈[0, 1] can be set to its maximum value (or near it), and gradually annealed towards zero. This has the effect of weighing the loss term from giving emphasis to the conv-gen operation, which is stable, to the conv-reg operation, which is closer to the desired behaviour of the KNet module.
To deal with the multicollinearity in our input space, we can use Tikhonov regularisation in our regression analysis. This ensures that the regression calculations are stable given any arbitrary input features. Contrast an ordinary least squares approach (linear regression analysis) with the Tikhonov regression analysis:
Wlinear=(ZTZ)−1ZTx (15.18)
WTikhonov=(ZTZ+λI)−1ZTx (15.19)
Here, Z is the design matrix (representing input features to the KNet module), x is the regression target (representing the target data, for example the ground truth image) and Wlinear and WTikhonov are the optimal weights produced from linear regression and Tikhonov regression, respectively.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2006275.8 | Apr 2020 | GB | national |
| 2008241.8 | Jun 2020 | GB | national |
| 2011176.1 | Jul 2020 | GB | national |
| 2012461.6 | Aug 2020 | GB | national |
| 2012462.4 | Aug 2020 | GB | national |
| 2012463.2 | Aug 2020 | GB | national |
| 2012465.7 | Aug 2020 | GB | national |
| 2012467.3 | Aug 2020 | GB | national |
| 2012468.1 | Aug 2020 | GB | national |
| 2012469.9 | Aug 2020 | GB | national |
| 2016824.1 | Oct 2020 | GB | national |
| 2019531.9 | Dec 2020 | GB | national |
This is a continuation of U.S. application Ser. No. 18/055,666, filed on Nov. 15, 2022, which is a continuation of U.S. application Ser. No. 17/740,716, filed on May 10, 2022, which is a continuation of International Application No. PCT/GB2021/051041, filed on Apr. 29, 2021, which claims priority to GB Application No. 2006275.8, filed on Apr. 29, 2020; GB Application No. 2008241.8, filed on Jun. 2, 2020; GB Application No. 2011176.1, filed on Jul. 20, 2020; GB Application No. 2012461.6, filed on Aug. 11, 2020; GB Application No. 2012462.4, filed on Aug. 11, 2020; GB Application No. 2012463.2, filed on Aug. 11, 2020; GB Application No. 2012465.7, filed on Aug. 11, 2020; GB Application No. 2012467.3, filed on Aug. 11, 2020; GB Application No. 2012468.1, filed on Aug. 11, 2020; GB Application No. 2012469.9, filed on Aug. 11, 2020; GB Application No. 2016824.1, filed on Oct. 23, 2020; GB Application No. 2019531.9, filed on Dec. 10, 2020; U.S. Provisional Application No. 63/017,295, filed on Apr. 29, 2020; and U.S. Provisional Application No. 63/053,807, filed Jul. 20, 2020, the entire contents of each of which being fully incorporated hereby by reference.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5048095 | Bhanu et al. | Sep 1991 | A |
| 9990687 | Kaufhold et al. | Jun 2018 | B1 |
| 10373300 | Besenbruch et al. | Aug 2019 | B1 |
| 10489936 | Zafar et al. | Nov 2019 | B1 |
| 10880551 | Topiwala et al. | Dec 2020 | B2 |
| 10886943 | Choi et al. | Jan 2021 | B2 |
| 10930263 | Mahyar | Feb 2021 | B1 |
| 10965948 | Appalaraju et al. | Mar 2021 | B1 |
| 11310509 | Topiwala et al. | Apr 2022 | B2 |
| 11330264 | Zhou et al. | May 2022 | B2 |
| 11375194 | Liu et al. | Jun 2022 | B2 |
| 11388416 | Habibian et al. | Jul 2022 | B2 |
| 11445222 | Andreopoulos et al. | Sep 2022 | B1 |
| 11481633 | Krishnamoorthy | Oct 2022 | B2 |
| 11526734 | Yang et al. | Dec 2022 | B2 |
| 11544536 | Gesmundo | Jan 2023 | B2 |
| 11610154 | Teig | Mar 2023 | B1 |
| 11748615 | Wu et al. | Sep 2023 | B1 |
| 20100332423 | Kapoor et al. | Dec 2010 | A1 |
| 20160292589 | Taylor et al. | Oct 2016 | A1 |
| 20170230675 | Wierstra | Aug 2017 | A1 |
| 20180139450 | Gao et al. | May 2018 | A1 |
| 20180176578 | Rippel et al. | Jun 2018 | A1 |
| 20190289296 | Kottke et al. | Sep 2019 | A1 |
| 20200021865 | Topiwala et al. | Jan 2020 | A1 |
| 20200027247 | Minnen et al. | Jan 2020 | A1 |
| 20200090069 | Mandt et al. | Mar 2020 | A1 |
| 20200097742 | Ratnesh Kumar et al. | Mar 2020 | A1 |
| 20200104640 | Poole et al. | Apr 2020 | A1 |
| 20200111501 | Sung et al. | Apr 2020 | A1 |
| 20200226421 | Almazan et al. | Jul 2020 | A1 |
| 20200304802 | Habibian et al. | Sep 2020 | A1 |
| 20200364574 | Kim et al. | Nov 2020 | A1 |
| 20200372686 | Wen et al. | Nov 2020 | A1 |
| 20200401916 | Rolfe et al. | Dec 2020 | A1 |
| 20210004677 | Menick et al. | Jan 2021 | A1 |
| 20210042606 | Bai et al. | Feb 2021 | A1 |
| 20210067808 | Schroers et al. | Mar 2021 | A1 |
| 20210142534 | Liu et al. | May 2021 | A1 |
| 20210152831 | Liu et al. | May 2021 | A1 |
| 20210166151 | Kennel et al. | Jun 2021 | A1 |
| 20210211741 | Andreopoulos et al. | Jul 2021 | A1 |
| 20210281867 | Golinski | Sep 2021 | A1 |
| 20210286270 | Middlebrooks et al. | Sep 2021 | A1 |
| 20210360259 | Wang et al. | Nov 2021 | A1 |
| 20210366161 | Wong | Nov 2021 | A1 |
| 20210390335 | Du et al. | Dec 2021 | A1 |
| 20210397895 | Sun et al. | Dec 2021 | A1 |
| 20220101106 | Van Der Wilk et al. | Mar 2022 | A1 |
| 20220103839 | Van Rozendaal et al. | Mar 2022 | A1 |
| 20230093734 | Zheng et al. | Mar 2023 | A1 |
| Entry |
|---|
| Balle et al. , “End-to-end optimized image compression,” arXiv preprint arXiv: 1611.01704 (2016). |
| Cheng et al. , “Energy compaction-based image compression using convolutional autoencoder,” IEEE Transactions on Multimedia 22.4, pp. 860-873 (2019). |
| Habibian, Amirhossein , et al., “Video Compression with Rate-Distortion Autoencoders,” arxiv.org, Cornell Univ. Library (Aug. 14, 2019) XP081531236. |
| Han, Jun , et al., “Deep Probabilistic Video Compression,” arxiv.org, Cornell Univ. Library, (Oct. 5, 2018) XP080930310. |
| Yan et al. , “Deep autoencoder-based lossy geometry compression for point clouds,” arXiv preprint arXiv: 1905.03691 (2019). |
| Chen , et al., “Neural ordinary differential equations,” Advances in neural information processing systems; 31 (2018). |
| Elsken , et al., “Neural architecture search: A survey,” The Journal of Machine Learning Research, 1997-2017 (2019). |
| Li , et al., “Sgas: Sequential Greedy Architecture Search,” In Proceedings of the IEEE/CVF Conf of Computer Vision and Pattern Recognition, pp. 1620-1630 (2020). |
| Molina , et al., “Pade Activation Units: End-to-end Learning of Flexible Activation Functions in Deep Networks,” arXiv preprint arXiv: 1907.06732 (2019). |
| Ziegler , et al., “Latent normalizing flows for discrete sequences,” Intl. Conf. on Machine Learning; PMLR (2019). |
| Leon-Garcia , “Probability and random processes for electrical engineering,” Pearson Education India (1994). |
| Number | Date | Country | |
|---|---|---|---|
| 20230388499 A1 | Nov 2023 | US |
| Number | Date | Country | |
|---|---|---|---|
| 63053807 | Jul 2020 | US | |
| 63017295 | Apr 2020 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 18055666 | Nov 2022 | US |
| Child | 18230240 | US | |
| Parent | 17740716 | May 2022 | US |
| Child | 18055666 | US | |
| Parent | PCT/GB2021/051041 | Apr 2021 | WO |
| Child | 17740716 | US |
