Patent Application
20170249536
Technical Field
The present invention is generally directed to the application of a Discrete Recurrent Neural Network (“DRNN”) to solving signal processing tasks. More specifically, the present invention provides a method, implementable in hardware or software, for learning a DRNN to signal domains for the application of solving various signal processing tasks of practical value in ways that significantly further the art.
Description of the Related Art
Processing of natural signals is a crucial component of most technology products. For example, media companies, including but not limited to social networking websites like Facebook, are taxed with hosting many high-resolution images and videos on their servers. Better image and video compression algorithms allow for reduced storage space in the aforementioned servers.
Signal domains have become increasingly more complex, and at the same time larger amounts of data are being sent through the internet, local networks, etc. Thus, it is important to have methods for compressing relevant data in real-time. Moreover, learning structure of data in the desired signal processing domain can and should be exploited for significant practical benefit.
There are several commercially available software and hardware modules for the compression of natural signals. Typically, these existing modules use linear coding to decompose a particular signal (“Image”) into a weighted sum of basis functions, e.g. the discrete cosine transform (DCT) basis. The basis functions are sometimes chosen without learning a dictionary from the input data, and the continuous-valued weighting coefficients must be quantized.
Several methods exist for learning internal configurations (“Network Parameters”) in neural network approaches to signal processing, including supervised deep learning methods, which use back propagation for training on labeled data. Other methods use Bayesian ideas (e.g. Maximum Likelihood) or sampling (e.g. Contrastive Divergence). Typically, approaches suffer from lack of scalability to high dimensions and often do not learn structure from data in an unsupervised manner.
Those skilled in the art will appreciate the benefits of a convex objective function with an easily computed gradient for high-dimensional discrete DRNN estimation (DRNN “Training” or “Learning”). There are no local minima that hinder training or necessitate random restarts with such an objective function. The tractability of the gradient herein contrasts with more standard procedures, such as maximum likelihood estimation, in that the number of terms to evaluate grows linearly in the number of network nodes, rather than exponentially.
In view of the foregoing, the present invention fulfills technical requirements for a general system, needed in the art, which can efficiently learn structure in a given high-dimensional signal domain in an iterative online fashion (“Online Learning”), and in real-time efficiently utilize this learned structure for feature-extraction, compression, data similarity measurement, clustering and classification, error-correcting coding, among other signal processing tasks of significant practical value.
The present invention is a method and system for efficiently extracting important features from data and processing them via a computer platform. The user inputs data from a particular signal domain into a platform, such as programming language PYTHON running on a Personal Computer (PC), or other implementations. The platform then trains a DRNN (the “DRNN Module”) on user-specified discretization of the data and uses network dynamics to identify states fixed under the action of the network dynamics (“Fixed-points” or “Attractors”) of the recurrent neural network, or memories (“Mem”s), for the input data. These memories are lossy perceptual features of input data.
The present invention as described herein implements the DRNN as a Hopfield network (i.e. symmetric McCulloch-Pitts network, auto-associative Perceptron, fully-observable Boltzmann machine) with a probabilistic model of binary data determined by a Lenz-Ising energy-based model. Other implementations are possible in which Hopfield-like networks are constructed such that the underlying probabilistic model includes higher-order correlations.
The present invention as described herein implements the Minimum Probability Flow (“MPF”) learning method for training the DRNN, but other training implementations are possible.
Typically, the DRNN operates on smaller decompositions of the signal (“Patches”), which themselves have characteristics (“Patch Statistics”). Depending on application, the user can adjust the discretization of the signal and other characteristics such as image patch specifications (e.g. length of sides for visual patches). When the input signal is a time-series such as an audio recording, the user can adjust the lengths of time-windows in patches and/or discretize the time-series in several possible ways.
Once the user selects a discretization of input data, the memory of any input data is returned as the output of the DRNN dynamics initialized at the discretization of the input. From this, the user can accumulate statistics on the distortion incurred by replacing the initial data with its corresponding memory (or its continuous replacement). These statistics can then be used to evaluate the coding cost of the Hopfield network relative to the achievable lower bound provided by the rate-distortion function. Thus, statistics output by the platform enable the user to validate memories as lossy perceptual features.
The user can also utilize the platform to compute the entropy rate of the initial data, the entropy rate of the corresponding fixed-points, and the number of memories as a function of the amount of input data. In various signal domains, the number of memories grows super-linearly with the size of the neural network, indicating that the incoming signals have sparse or other structure that increases the memory capacity of these networks. The dynamics used to obtain the memory or feature corresponding to an initial condition often converge very quickly, so that this neural network provides an efficient “look-up table” which maps the original data to memories of the DRNN.
Existing approaches to identifying lossy perceptual features often rely on linear coding, in which the data is expressed as a truncated sum of weighted basis functions. In the present invention, these basis functions are replaced by memories of the DRNN trained on data. The user's discretization scheme, when properly chosen, endows these memories with sufficient expressive power to avoid standard quantization algorithms. Thus, the memories of the DRNN with appropriate discretization have reduced coding cost relative to other standard methods for compressing and modeling signals.
The user can access different levels of loss (“Distortion”) in compression by tuning various hyperparameters, including but not limited to those mentioned earlier, related to the discretization scheme, and thus interpolate between storing features producing high distortion and those that are nearly lossless. In other words, users can choose the acceptable level of perceptual distortion, and the DRNN Module can achieve that level by careful setting of hyperparameters.
The present invention allows for a method to extract information out of signal domains and apply such knowledge to solve specific signal processing tasks of significant practical value.
Although the exemplary embodiments will be generally described in the context of software modules running in a distributed computing environment, those skilled in the art will recognize that the present invention also can be implemented in conjunction with other program modules for other types of computers. In a distributed computing environment, program modules may be physically located in different local and remote memory storage devices. Execution of the program modules may occur locally in a stan-alone manner or remotely in a client/server manner. Examples of such distributed computing environments include local area networks of an office, enterprise-wide computer networks, and the global internet.
The detailed description that follows is represented largely in terms of processes and symbolic representations of operations in a distributed computing environment by conventional computer components, including remote file servers, remote computer servers, remote memory storage devices, a central processing unit (CPU), memory storage devices for the CPU, display devices and input devices. Each of these conventional distributed computing components is accessible by the CPU via a communications network.
The processes and operations performed by the computer include the manipulation of signals by a CPU or remote server and the maintenance of these signals within data structures resident in one or more of the local or remote memory storage devices. Such data structures impose a physical organization upon the collection of data stored within a memory storage device and represent specific electrical or magnetic elements. These symbolic representations are the means used by those skilled in the art of computer programming and computer construction to most effectively convey teachings and discoveries to others skilled in the art.
The present invention also includes implementation of a computer program, which embodies the functions described herein and illustrated in the appended flow charts. However, it should be apparent that there could be many different ways of implementing the invention in computer programming, and the invention should not be construed as limited to any one set of computer program instructions. Further, a skilled programmer would be able to write such a computer program to implement the disclosed invention without difficulty based on the flow charts and associated description in the application text, for example. Therefore, disclosure of a particular set of program code instructions is not considered necessary for an adequate understanding of how to make and use the invention. The inventive functionality of the claimed computer program will be explained in more detail in the following description in conjunction with the remaining figures illustrating the program flow.
Referring now to the drawings, in which like numerals represent like elements throughout the several figures, aspects of the present invention and the preferred operating environment will be described.
In one exemplary embodiment, we illustrate application 140 to signal domain 110. However, as noted above, the invention should not be construed as limited to any one application or domain. Further, a skilled programmer would be able to write such a module to implement the disclosed invention without difficulty based on the flow charts, and associated description in the application text, for example. Therefore, disclosure of a particular set of program code instructions, such as in programming language PYTHON, is not considered necessary for an adequate understanding of how to make and use the present invention.
Referring now to
Referring to FIGS. 6ABC, application of present invention to unsupervised clustering 150 is illustrated, demonstrating discovery of recurrent patterns of activity in the exemplary signal domain of electrophysiological recordings of neurons.
Referring now to
Referring to
In conclusion, the present invention enables several applications of a DRNN Module trained to extract information of practical value from several signal domains.
It will be appreciated that the present invention fulfills the needs of the prior art described herein and meets the above-stated objects. While there has been shown and described the preferred embodiment of the invention, it will be evident to those skilled in the art that various modifications and changes may be made thereto without departing from the spirit and the scope of the invention set forth in the appended claims and equivalence thereof. Although the present invention has been described as operating on the specific signal domains described herein, it should be understood that the invention can be applied to other types of signal domains such as audio, video, user meta-data, internet traffic, etc.
This application claims priority to provisional patent application entitled “Self-organizing discrete recurrent network digital image codec,” filed on Feb. 27, 2016 and assigned U.S. Provisional Application No. 62/126,141.
This invention was made with government support under contract 0917342 with IIS Division of Information & Intelligent Systems awarded by the National Science Foundation. The government has certain rights in the invention.
