Patent Application
20240220814
The present invention relates to a learning device, a learning method, and a learning program.
Conventionally, a deep generation model that is a technique based on a deep learning technique and generates a sample close to a real thing by learning a distribution of learned data is known. For example, a generative adversarial network (GAN) is known as a deep generation model (e.g., refer to Non Patent Literature 1).
Moreover, for example, variational auto encoders (VAEs) (Reference Literature 1: Kingma, Diederik P., and Max Welling. “Auto-encoding variational bayes.” arXiv preprint arXiv: 1312.6114 (2013). (ICLR 2014)) are known as other deep generation models.
However, conventional techniques have a problem that over-learning may occur and the accuracy of the model may not be improved. For example, a high frequency component not included in actual learning data is mixed into a sample generated by a generator of a learned GAN. As a result, a discriminator performs authenticity determination depending on a high frequency component, and over-learning may occur.
In order to solve the above-described problem and achieve an object, a learning device includes: a removal unit configured to remove a predetermined component from a frequency component obtained by transforming data of a predetermined domain; a calculation unit configured to calculate a loss function on the basis of a result obtained by inputting data obtained by returning the frequency component, from which the predetermined component has been removed by the removal unit, to the predetermined domain into a discriminator constituting an adversarial learning model; and an update unit configured to update a parameter of the adversarial learning model so that the loss function is optimized.
According to the present invention, the occurrence of over-learning can be suppressed, and the accuracy of the model can be improved.
Hereinafter, embodiments of a learning device, a learning method, and a learning program according to the present application will be described in detail with reference to the drawings. Note that the present invention is not limited to the embodiments described below.
A GAN is a technique of learning a data distribution p_data(x) using two deep learning models of a generator G and a discriminator D. The G learns to deceive the D, and the D learns to distinguish the G from learning data. A model in which a plurality of such models has an adversarial relationship may be referred to as an adversarial learning model.
An adversarial learning model such as the GAN is used in generation of images, texts, voices, and the like.
Here, the GAN has a problem that the D over-learns a learning sample as the learning progresses. As a result, each model cannot perform meaningful update to data generation, and the generation quality of the generator deteriorates.
Moreover, Reference Literature 2 describes that a learned CNN output performs prediction depending on a high frequency component of an input.
Reference Literature 2: Wang, Haohan, et al. “High-frequency Component Helps Explain the Generalization of Convolutional Neural Networks.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. (CVPR 2020)
Therefore, an object of the first embodiment is to suppress the occurrence of over-learning and improve the accuracy of the model by removing the high frequency component of the data inputted into the discriminator D.
As illustrated in
Reference Literature 3: Durall, Ricard, Margret Keuper, and Janis Keuper. “Watch your Up-Convolution: CNN Based Generative Deep Neural Networks are Failing to Reproduce Spectral Distributions.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. (CVPR 2020)
Returning to
In the GAN, the discriminator D is optimized so that the discrimination accuracy of the discriminator D is improved, that is, the probability that the discriminator D discriminates Real as Real increases. Moreover, the generator G is optimized so that the ability of the generator G to deceive the generator G, that is, the probability that the discriminator D discriminates Real as Fake increases.
In addition to the above optimization, the generator G is optimized so that frequency components of Real and Fake match in the present embodiments. The following description will explain details of learning processing of the deep learning model in addition to a configuration of a learning device of the embodiments.
The input/output unit 11 is an interface for inputting/outputting data. For example, the input/output unit 11 may be a communication interface such as a network interface card (NIC) for performing data communication with another device via a network. Moreover, the input/output unit 11 may be an interface to be connected with an input device such as a mouse or a keyboard, and an output device such as a display.
The storage unit 12 is a storage device such as a hard disk drive (HDD), a solid state drive (SSD), or an optical disk. Note that the storage unit 12 may be a data-rewritable semiconductor memory such as a random access memory (RAM), a flash memory, or a non volatile static random access memory (NVSRAM). The storage unit 12 stores an operating system (OS) and various programs to be executed by the learning device 10. Moreover, the storage unit 12 stores model information 121.
The model information 121 is information such as parameters for constructing a deep learning model, and is appropriately updated in the learning processing. Moreover, the updated model information 121 may be outputted to another device or the like via the input/output unit 11.
The control unit 13 controls the entire learning device 10. The control unit 13 is, for example, an electronic circuit such as a central processing unit (CPU), a micro processing unit (MPU), or a graphics processing unit (GPU), or an integrated circuit such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA). Moreover, the control unit 13 has an internal memory for storing programs defining various processing procedures and control data, and executes each processing using the internal memory. Moreover, the control unit 13 functions as various processing units by operation of various programs. For example, the control unit 13 includes a generation unit 131, a transformation unit 132, a removal unit 133, a calculation unit 134, and an update unit 135.
The generation unit 131 inputs the random number z into the generator G to generate data.
The transformation unit 132 transforms data inputted into the discriminator D into frequency components. The transformation unit 132 transforms real data (Real) and data (Fake) generated by the generator into frequency components.
The removal unit 133 removes a predetermined component from a frequency component obtained by transforming data of a predetermined domain. In the embodiment, the removal unit 133 removes a high frequency component.
Here, processing of removing a high frequency component by the transformation unit 132 and the removal unit 133 will be described with reference to
As illustrated in
The xreal is real data, and is referred to as first data herein. Moreover, the xfake is data generated by the generator, and is referred to as second data herein. Moreover, the transformation unit 132 transforms the first data into a first frequency component and transforms the second data into a second frequency component.
Next, the removal unit 133 removes (filtering, masking) the high frequency component by Formula (1) in F-Drop. The x is one of the first data xreal and the second data xfake. F(⋅) is a function for performing frequency transformation by DFT and DCT.
However, each component of the function M is calculated by Formula (2).
Here, each data in the frequency space (frequency domain) is represented by coordinates on the u axis and the v axis. Moreover, the right side of the inequality of Formula (2) is determined according to the data.
For example, in a case where the first data and the second data are image data, H is a height of the image, and W is a width of the image. The height H and the width W are represented by, for example, the number of pixels. Moreover, in this case, the image data before transformation is data in an RGB space in which each element is represented by an RGB value.
Here, when the size of the image data is 5×5, the components illustrated in
The numerical value in each cell in
It can be said that (u2+v2)1/2 indicates a distance from the origin in the frequency space. Therefore, the removal unit 133 removes a component whose distance from the origin in the frequency domain is equal to or longer than a threshold from the first frequency component obtained by transforming the first image data in the RGB space and from the second frequency component obtained by transforming the second image data in the RGB space generated by the generator constituting the adversarial learning model.
Furthermore, the transformation unit 132 returns the data, from which a component has been removed by the removal unit 133, to the space before transformation. For example, when transformation by discrete cosine transform (DCT) is performed, the transformation unit 132 performs inverse transformation by inverse discrete cosine transform (IDCT).
In a case where the original data is image data in the RGB space, the transformation unit 132 transforms data in the frequency space, from which a high frequency component of Formula (1) has been removed, into data in the RGB space by inverse transformation.
In this manner, the removal unit 133 removes a predetermined component from the first frequency component obtained by transforming the first data and from the second frequency component obtained by transforming the second data generated by the generator constituting the adversarial learning model.
The calculation unit 134 calculates a loss function on the basis of a result obtained by inputting data obtained by returning the frequency component, from which the predetermined component has been removed by the removal unit 133, to a predetermined domain into the discriminator constituting the adversarial learning model.
The calculation unit 134 calculates a loss function that becomes larger as the discrimination accuracy of the discriminator becomes lower for each of data whose first frequency component and second frequency component, from which the predetermined component has been removed, are returned to the predetermined domain.
The update unit 135 updates a parameter of the adversarial learning model so that the loss function is optimized. For example, the update unit 135 updates a parameter of the generator so that the loss function is optimized.
For example, the calculation unit 134 and the update unit 135 update the parameter using a loss function used in a known adversarial learning model (GAN).
Next, the learning device 10 samples a random number z from a normal distribution, and creates a sample (Fake) by G(z) (step S102).
Here, the learning device 10 calculates Drop(Real, γ) and Drop(Fake, γ), and inputs the results into the discriminator D (step S103). The function Drop(⋅) is as described in Formula (1).
Here, the learning device 10 calculates a GAN loss function of the generator G (step S104).
Furthermore, the learning device 10 updates the parameter of the generator G by the back error propagation method of the overall loss (here, a GAN loss function) (step S105).
Moreover, the learning device 10 learns the discriminator D (step S106).
At this time, in a case where the maximum number of learning steps>the number of learning steps is satisfied (step S107, True), the learning device 10 returns to step S101 and repeats the processing. On the other hand, in a case where the maximum number of learning steps>the number of learning steps is not satisfied (step S107, False), the learning device 10 terminates the processing.
As described above, the removal unit 133 removes a predetermined component from the frequency component obtained by transforming the data of the predetermined domain. The calculation unit 134 calculates a loss function on the basis of a result obtained by inputting data obtained by returning the frequency component, from which the predetermined component has been removed by the removal unit 133, to a predetermined domain into the discriminator constituting the adversarial learning model. The update unit 135 updates a parameter of the adversarial learning model so that the loss function is optimized.
As described above, the generator G and the discriminator D in the GAN may be excessively concentrated on the high frequency component of the data, and over-learning may occur. For example, when the discriminator D performs authenticity determination depending on the high frequency component, the generator G learns the high frequency component to deceive the discriminator D. Then, the result of the authenticity determination depends only on the high frequency component, and effective update for bringing the data distribution closer is not performed.
On the other hand, the learning device 10 can remove the high frequency component (frequency drop) and perform learning of the GAN after removing the high frequency component.
As a result, according to the present embodiment, it is possible to suppress the deviation (frequency gap) of the frequency component from the learning data generated in learning of the GAN. Furthermore, the data generation quality of the generator G is also improved as the property in the frequency component becomes closer.
As described above, according to the present embodiment, the occurrence of over-learning can be suppressed, and the accuracy of the model can be improved.
The removal unit 133 removes a predetermined component from the first frequency component obtained by transforming the first data and from the second frequency component obtained by transforming the second data generated by the generator constituting the adversarial learning model. The calculation unit 134 calculates a loss function that becomes larger as the discrimination accuracy of the discriminator becomes lower for each of data whose first frequency component and second frequency component, from which the predetermined component has been removed, are returned to the predetermined domain. The update unit 135 updates the parameter of the generator so that the loss function is optimized.
As described above, it is possible to further improve the accuracy of the model by removing the high frequency component from both the real data and the generated data in the GAN.
The removal unit 133 removes a component whose distance from the origin in the frequency domain is equal to or longer than a threshold from the first frequency component obtained by transforming the first image data in the RGB space and from the second frequency component obtained by transforming the second image data in the RGB space generated by the generator constituting the adversarial learning model.
As a result, according to the embodiment, the high frequency component can be removed from the image data.
The learning device 10 may cause the loss function to include the frequency component matching loss of the generator G and the discriminator D. In the second embodiment, the learning device 10 optimizes the frequency component matching loss at the time of learning.
The processing of the generation unit 131 and the transformation unit 132 is similar to that of the first embodiment.
The calculation unit 134 further calculates an inter-data error between the first frequency component and the second frequency component. The calculation unit 134 can calculate the error by an arbitrary method such as the mean square error (MSE), the root mean square error (RMSE), or L1. Here, the calculation unit 134 calculates LD in Formula (3) and LG in Formula (4). Moreover, the calculation unit 134 calculates the inter-data error Lfreq (frequency component matching loss) by Formula (5).
Here, Xreal and Xfake are batches of Real and Fake, respectively. Moreover, |Xreal| and |Xfake| are batch sizes thereof, respectively. Real is real data. Moreover, Fake is data generated by the generator G.
Moreover, F(⋅) is a function that transforms data in a spatial domain into a frequency component. The xreali and the xfakej are i-th data of Xreal and j-th data of Xfake, respectively, and are examples of the first data and the second data. Moreover, F(xreali) corresponds to the first frequency component. Moreover, F(xfakej) corresponds to the second frequency component.
In this manner, the calculation unit 134 calculates an error between a batch average of the plurality of first frequency components obtained by transforming the plurality of first data and a batch average of the plurality of second frequency components obtained by transforming the plurality of second data. That is, the error here corresponds not to an error between single data samples but to an error between batch averages.
Furthermore, the calculation unit 134 calculates a loss function LG that becomes larger as the error between the first frequency component and the second frequency component becomes larger and that becomes larger as the accuracy of discrimination of the first data and the second data by the discriminator constituting the adversarial learning model becomes lower as in Formula (4). The λ is a hyperparameter that functions as a weight.
The G(⋅) is a function that outputs data (Fake) generated by the generator G on the basis of an argument. Moreover, D(⋅) is a function that outputs a probability that the discriminator D discriminates data inputted as an argument as Real.
The update unit 135 updates the parameter of the adversarial learning model so that both the loss function and the inter-data error are optimized. Specifically, the update unit 135 updates the parameter of the generator G so that the loss function LG in Formula (4) is optimized.
Moreover, the update unit 135 updates the parameter of the discriminator D so that the loss function LD of Formula (3) is optimized. Here, x is real data (Real).
Next, the learning device 10 samples a random number z from a normal distribution, and generates a sample (Fake) by G(z) (step S202). Moreover, the learning device 10 transforms Real and Fake into frequency components by DCT or DFT, and then calculates the batch average of the frequency components (step S203).
Here, the learning device 10 calculates Drop(Real, γ) and Drop(Fake, γ), and inputs the results into the discriminator D (step S204). The function Drop(⋅) is as described in Formula (1).
The learning device 10 calculates a GAN loss function of the generator G (step S205). The GAN loss of the generator G corresponds to a first term on the right side of Formula (4). Then, the learning device 10 calculates a frequency component matching loss from the batch average of the Real-Fake frequency components (step S206). The frequency component matching loss corresponds to Lfreq in Formula (5).
Furthermore, the learning device 10 calculates the sum of the GAN loss function regarding G and the frequency component matching loss as an overall loss (step S207). The overall loss corresponds to LG in Formula (4). The learning device 10 may multiply the frequency component matching loss by weight λ. The learning device 10 updates the parameter of the generator G by the back error propagation method of the overall loss (step S208).
Moreover, the learning device 10 learns the discriminator D (step S209). Specifically, the learning device 10 updates the parameter of the discriminator D by the back error propagation method of the loss function LD of Formula (3).
At this time, in a case where the maximum number of learning steps>the number of learning steps is satisfied (Step S210, True), the learning device 10 returns to step S101 and repeats the processing. On the other hand, in a case where the maximum number of learning steps>the number of learning steps is not satisfied (Step S210, False), the learning device 10 terminates the processing.
The calculation unit 134 further calculates an inter-data error between the first frequency component and the second frequency component. The update unit 135 updates the parameter of the adversarial learning model so that both the loss function and the inter-data error are optimized.
As a result, the influence of the frequency component in learning of the adversarial learning model can be further reduced.
[Experiment]
An experiment for actually carrying out the above embodiment will be described. The experimental settings are as follows.
In this way, the high frequency component does not influence the appearance of an image for humans. This is because natural images recognized by humans are concentrated on low frequency components.
Moreover, each component of each illustrated device is functionally conceptual, and does not necessarily need to be physically configured as illustrated. That is, a specific form of distribution and integration of devices is not limited to the illustrated form, and all or some thereof can be functionally or physically distributed or integrated in an arbitrary unit according to various loads, usage conditions, and the like. The whole or an arbitrary part of each processing function performed in each device can be implemented by a central processing unit (CPU) and a program analyzed and executed by the CPU, or may be implemented as hardware by wired logic. Moreover, the program may be executed not only by a CPU but also by another processor such as a GPU.
Moreover, among the processes described in the present embodiment, all or some of the processes described as being automatically performed can be manually performed, or all or some of the processes described as being manually performed can be automatically performed by a known method. In addition, the processing procedure, the control procedure, specific names, and information including various data and parameters illustrated in the literatures and the drawings can be arbitrarily changed unless otherwise specified.
As an embodiment, the learning device 10 can be mounted by installing a learning program for executing the above learning processing as package software or online software in a desired computer. For example, it is possible to cause an information processing device to function as the learning device 10 by causing the information processing device to execute the above learning program. The information processing device mentioned here includes a desktop or notebook personal computer. In addition, the information processing device includes a mobile communication terminal such as a smartphone, a mobile phone, and a personal handyphone system (PHS), a slate terminal such as a personal digital assistant (PDA), and the like.
Moreover, the learning device 10 can also be mounted as a learning server device that uses a terminal device used by the user as a client and provides the client with a service related to the learning processing described above. For example, the learning server device is mounted as a server device that provides a learning service having learning data as an input and information of a learned model as an output. In this case, the learning server device may be mounted as a web server, or may be mounted as a cloud that provides a service related to the learning processing by outsourcing.
The memory 1010 includes a read only memory (ROM) 1011 and a random access memory (RAM) 1012. The ROM 1011 stores, for example, a boot program such as a basic input output system (BIOS). The hard disk drive interface 1030 is connected with a hard disk drive 1090. The disk drive interface 1040 is connected with a disk drive 1100. For example, a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected with, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected with, for example, a display 1130.
The hard disk drive 1090 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, the program that defines each processing of the learning device 10 is mounted as the program module 1093 in which codes executable by a computer are described. The program module 1093 is stored in, for example, the hard disk drive 1090. For example, the program module 1093 for executing processing similar to the functional configuration in the learning device 10 is stored in the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced with a solid state drive (SSD).
Moreover, the setting data to be used in the processing of the above-described embodiment is stored in, for example, the memory 1010 or the hard disk drive 1090 as the program data 1094. Then, the CPU 1020 reads out the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1090 to the RAM 1012 as necessary, and executes the processing of the above-described embodiment.
Note that the program module 1093 and the program data 1094 are not limited to being stored in the hard disk drive 1090, and may be stored in, for example, a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and the program data 1094 may be stored in another computer connected via a network (a local area network (LAN), a wide area network (WAN), or the like). Then, the program module 1093 and the program data 1094 may be read by the CPU 1020 from another computer via the network interface 1070.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2021/020306 | 5/27/2021 | WO |
