Data processing method, information processing device, and program

Abstract

The present invention improves a technology related to data processing using a neural network model. The present invention provides a data processing method executed by an information processing device. The data processing method includes a step for training each of a first neural network model, a second neural network model, and a third neural network model using first data, second data, and third data as teacher data so that a first weight coefficient related to the first neural network model, a second weight coefficient related to the second neural network model, and a third weight coefficient related to the third neural network model satisfy a first constraint condition. The first constraint condition is a condition for performing training so as to express the third data on the basis of the first weight coefficient and the second weight coefficient.

Description

Data processing method, information processing device, and program CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority from Japanese Patent Application No. 2024-004188, filed on January 15, 2024, the entire disclosure of which is incorporated herein by reference.

[0002] The present disclosure relates to a data processing method, an information processing device, and a program.

[0003] Conventionally, data processing and conversion techniques using neural network models such as Implicit Neural Representations (INR) have been known (for example, Non-Patent Document 1).

[0004] Vincent Sitzmann et al. "Implicit Neural Representations with Periodic Activation Functions" arXiv (2000)

[0005] Even when neural network models learn the same thing, there is a high degree of redundancy, as various combination patterns of weight coefficients are obtained as a result of learning depending on the learning process, the initial value settings of the weight coefficients, etc. However, in conventional INR, utilizing this redundancy has never been considered. On the other hand, utilizing redundancy could be considered, for example, to create a certain relationship between the weight coefficients of multiple learning models. As such, there is room for improvement in technology related to data processing and conversion using neural network models.

[0006] In view of the above circumstances, an object of the present disclosure is to improve the technology relating to data processing and conversion using neural network models.

[0007] (1) A data representation method according to one embodiment of the present disclosure is a data processing method executed by an information processing device, comprising: using first data, second data, and third data as training data to train a first neural network model, a second neural network model, and a third neural network model, respectively, so that a first weighting coefficient associated with the first neural network model, a second weighting coefficient associated with the second neural network model, and a third weighting coefficient associated with the third neural network model satisfy a first constraint condition, wherein the first constraint condition is a condition for training to represent the third data based on the first weighting coefficient and the second weighting coefficient.

[0008] (2) A data processing method according to an embodiment of the present disclosure is the data processing method according to (1), in which the third weighting factor is an average value of the first weighting factor and the second weighting factor.

[0009] (3) A data processing method according to an embodiment of the present disclosure is the data processing method according to (1), in which the third weighting factor is a weighted average value of the first weighting factor and the second weighting factor.

[0010] (4) A data processing method according to one embodiment of the present disclosure is the data processing method described in any one of (1) to (3), further including: using fourth data as training data to train a fourth neural network model so that the first weighting coefficient, the second weighting coefficient, and the fourth weighting coefficient related to the fourth neural network model satisfy a second constraint condition, wherein the second constraint condition is a condition for training to represent the fourth data based on the first weighting coefficient and the second weighting coefficient.

[0011] (5) A data processing method according to an embodiment of the present disclosure is the data processing method according to any one of (1) to (3), wherein activation functions of the first neural network model, the second neural network model, and the third neural network model are sine functions.

[0012] (6) A data processing method according to one embodiment of the present disclosure is a data representation method according to any one of (1) to (5), wherein in the learning step, the initial value of the first weighting coefficient and the initial value of the second weighting coefficient are the same.

[0013] (7) A data processing method according to an embodiment of the present disclosure is the data processing method described in any one of (1) to (6), wherein the number of layers of the first neural network model, the number of layers of the second neural network model, and the number of layers of the third neural network model are the same.

[0014] (8) A data processing method according to an embodiment of the present disclosure is a data processing method executed by an information processing device, and includes the steps of: using first data, second data, and third data as training data to respectively train a first neural network model, a second neural network model, and a third neural network model so that a first weight coefficient associated with the first neural network model, a second weight coefficient associated with the second neural network model, and a third weight coefficient associated with the third neural network model satisfy a first constraint condition, and acquiring the first weight coefficient and the second weight coefficient such that the first constraint condition is a condition for training to represent the third data based on the first weight coefficient and the second weight coefficient; generating the third weight coefficient based on the first weight coefficient and the second weight coefficient; and processing data based on the third weight coefficient.

[0015] (9) An information processing device according to an embodiment of the present disclosure is an information processing device including a control unit, wherein the control unit uses first data, second data, and third data as training data to train a first neural network model, a second neural network model, and a third neural network model, respectively, so that a first weighting coefficient associated with the first neural network model, a second weighting coefficient associated with the second neural network model, and a third weighting coefficient associated with the third neural network model satisfy a first constraint condition, and the first constraint condition is a condition for training to represent the third data based on the first weighting coefficient and the second weighting coefficient.

[0016] (10) A program according to an embodiment of the present disclosure causes a computer to execute the following: using first data, second data, and third data as training data to train a first neural network model, a second neural network model, and a third neural network model, respectively, so that a first weighting coefficient associated with the first neural network model, a second weighting coefficient associated with the second neural network model, and a third weighting coefficient associated with the third neural network model satisfy a first constraint condition, wherein the first constraint condition is a condition for training to represent the third data based on the first weighting coefficient and the second weighting coefficient.

[0017] According to one embodiment of the present disclosure, techniques for data processing and transformation using neural network models are improved.

[0018] 1 is a block diagram showing a schematic configuration of a system according to an embodiment of the present disclosure. FIG. 2 is a block diagram showing a schematic configuration of an information processing device. FIG. 3 is a flowchart showing the operation of the information processing device 10a. FIG. 4 is a schematic configuration of a mixed model including a first neural network model, a second neural network model, and a third neural network model. FIG. 5 is a conceptual diagram showing a learning process of the mixed model of FIG. 4. FIG. 6 is a schematic diagram showing a mixed model in which a first constraint condition is that a third weighting coefficient is the average value of a first weighting coefficient and a second weighting coefficient. FIG. 7 is a flowchart showing the operation of the information processing device 10b. FIG. 8 is a schematic configuration of a mixed model including a first neural network model, a second neural network model, a third neural network model, and a fourth neural network model. FIG. 9 is a diagram showing a case in which, in the mixed model of FIG. 8, images of each training data are partial images obtained by dividing an original image into four parts.

[0019] Hereinafter, embodiments of the present disclosure will be described.

[0020] (Outline of the embodiment) The outline and configuration of a system 1 according to the present embodiment will be described with reference to FIG.

[0021] The system 1 according to this embodiment includes an information processing device 10a and an information processing device 10b. The information processing device 10a and the information processing device 10b are communicatively connected to a network 20, which may include, for example, a mobile communication network and the Internet. The information processing device 10a and the information processing device 10b do not need to be communicatively connected at all times, and data may be transmitted between them via an information storage medium such as a memory. In this specification, when the information processing device 10a and the information processing device 10b are not distinguished from each other, they will be collectively referred to as the information processing device 10.

[0022] The information processing device 10 is any device used by a user. A general-purpose computer or a dedicated computer can be used as the information processing device 10. Although FIG. 1 shows an example in which the system 1 includes one information processing device 10, the present invention is not limited to this. The system 1 may include two or more information processing devices 10.

[0023] First, an overview of the present technology will be described for a case where the information processing device 10a and the information processing device 10b function as an encoding device and a decoding device, respectively. For example, the information processing device 10a encodes data such as an image, and the information processing device 10b decodes data such as text, images, and audio from the encoded signal. Here, in this embodiment, the image includes a still image and a moving image.

[0024] In the encoding and decoding processes according to this embodiment, a data representation technique using a neural network model is used. For example, the information processing device 10a uses image data as training data to train the neural network model 30a. In FIG. 1, an image of 256×256 size is used as training data as an example. The coordinates (x, y) of each pixel are normalized to the range of [-1, 1], for example, to obtain a value (p ｘ , p ｙ ) is input to the neural network model 30a. The weighting coefficient θ of the neural network model 30a １ (θ １ includes the weights themselves and bias b1, and refers to the entire set of parameters for the neural network. (The same applies below) is updated based on the learning process to output the pixel values ​​(RGB) of the image that is the training data.

[0025] The information processing device 10b uses a weighting coefficient θ １ In other words, the image data is decoded using the weighting coefficient θ １ The weighting coefficient θ １ The data is transmitted from the information processing device 10a to the information processing device 10b via the network 20. In FIG. 1, an example is shown in which the data is enlarged to an image of 1024×1024 size and decoded. The coordinates (x, y) of each pixel are normalized to the range of [−1, 1] (p ｘ , p ｙ ) is input to the neural network model 30b. １By using the above, the image data of the teacher data used for learning by the information processing device 10a is enlarged and decoded. Note that although an example in which the information processing device 10b enlarges and decodes the image data has been described here, the present invention is not limited to this. The information processing device 10b may decode an image of the same size (256 x 256) as the original image data, or may reduce the size.

[0026] The system 1 according to this embodiment is an improved version of the above-described technology, which uses a neural network model to represent data. In summary, the information processing device 10 uses first data, second data, and third data as training data to train a first neural network model, a second neural network model, and a third neural network model, respectively, so that the first weighting coefficient of the first neural network model, the second weighting coefficient of the second neural network model, and the third weighting coefficient of the third neural network model satisfy a first constraint. Hereinafter, in this embodiment, the entire training model including the first neural network model, the second neural network model, and the third neural network model is also referred to as a mixed model. Here, the first constraint is a condition under which the third data is trained to be represented based on the first and second weighting coefficients. In other words, the information processing device 10 according to this embodiment represents the third data based on the first and second weighting coefficients.

[0027] As described above, according to this embodiment, the first neural network model, the second neural network model, and the third neural network model are trained so that the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient satisfy the first constraint. That is, in this embodiment, redundancy in the neural network models is utilized to impart a predetermined relationship to the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient. This allows the third data to be represented based on the first weighting coefficient and the second weighting coefficient. Therefore, for example, the third data can be decoded by transmitting only the first weighting coefficient and the second weighting coefficient to the decoding-side information processing device, which is advantageous in terms of improving data transmission efficiency, etc. In other words, technology related to data representation using neural network models is improved in that the third data can be decoded by transmitting the weighting coefficients of the first data and the weighting coefficients of the second data.

[0028] (Configuration of Information Processing Apparatus) As shown in FIG. 2 , the information processing apparatus 10 includes a control unit 11 , a storage unit 12 , an input unit 13 , an output unit 14 , and a communication unit 15 .

[0029] The control unit 11 includes at least one processor, at least one dedicated circuit, or a combination thereof. The processor is a general-purpose processor such as a central processing unit (CPU) or a graphics processing unit (GPU), or a dedicated processor specialized for a specific process. The dedicated circuit is, for example, a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC). The control unit 11 controls each unit of the information processing device 10 and executes processes related to the operation of the information processing device 10.

[0030] The storage unit 12 includes at least one semiconductor memory, at least one magnetic memory, at least one optical memory, or a combination of at least two of these. The semiconductor memory is, for example, a random access memory (RAM) or a read-only memory (ROM). The RAM is, for example, a static random access memory (SRAM) or a dynamic random access memory (DRAM). The ROM is, for example, an electrically erasable programmable read-only memory (EEPROM). The storage unit 12 functions as, for example, a main storage device, an auxiliary storage device, or a cache memory. The storage unit 12 stores data used in the operation of the information processing device 10 and data obtained by the operation of the information processing device 10.

[0031] The input unit 13 includes at least one input interface. The input interface may be, for example, a physical key, a capacitance key, a pointing device, or a touch screen integrated with a display. The input interface may also be, for example, a sound sensor that accepts voice input, or a camera that accepts gesture input. The input unit 13 accepts an operation to input data used in the operation of the information processing device 10. The input unit 13 may be connected to the information processing device 10 as an external input device instead of being provided in the information processing device 10. Any connection method may be used, for example, a Universal Serial Bus (USB), a High-Definition Multimedia Interface (HDMI) (registered trademark), or Bluetooth (registered trademark).

[0032] The output unit 14 includes at least one output interface. The output interface is, for example, a display that outputs information as video, or a speaker that outputs information as audio. The display is, for example, an LCD (liquid crystal display) or an organic EL (electro luminescence) display. The output unit 14 displays and outputs data obtained by the operation of the information processing device 10. The output unit 14 may be connected to the information processing device 10 as an external output device instead of being provided in the information processing device 10. Any connection method can be used, for example, USB, HDMI (registered trademark), or Bluetooth (registered trademark).

[0033] The communication unit 15 includes at least one external communication interface. The communication interface may be either a wired or wireless communication interface. In the case of wired communication, the communication interface is, for example, a LAN (Local Area Network) interface or a USB (Universal Serial Bus). In the case of wireless communication, the communication interface is, for example, an interface compatible with mobile communication standards such as LTE (Long Term Evolution), 4G (4th generation), or 5G (5th generation), or an interface compatible with short-range wireless communication such as Bluetooth (registered trademark). The communication unit 15 receives data used in the operation of the information processing device 10 and transmits data obtained by the operation of the information processing device 10.

[0034] The functions of the information processing device 10 are realized by executing a program according to this embodiment on a processor corresponding to the information processing device 10. That is, the functions of the information processing device 10 are realized by software. The program causes a computer to execute the operations of the information processing device 10, thereby causing the computer to function as the information processing device 10. That is, the computer functions as the information processing device 10 by executing the operations of the information processing device 10 in accordance with the program.

[0035] In this embodiment, the program can be recorded on a computer-readable recording medium. The computer-readable recording medium includes non-transitory computer-readable media, such as a magnetic recording device, an optical disc, a magneto-optical recording medium, or a semiconductor memory. The program can be distributed, for example, by selling, transferring, or lending a portable recording medium, such as a DVD (digital versatile disc) or a CD-ROM (compact disc read only memory), on which the program is recorded. The program can also be distributed by storing the program in the storage of an external server and transmitting the program from the external server to another computer. The program can also be provided as a program product.

[0036] Some or all of the functions of the information processing device 10 may be implemented by a dedicated circuit equivalent to the control unit 11. In other words, some or all of the functions of the information processing device 10 may be implemented by hardware.

[0037] (Operation of Information Processing Apparatus) The operation of the information processing apparatus 10 according to this embodiment will be described with reference to Fig. 3 and Fig. 7. First, the steps of encoding an image in the information processing apparatus 10a will be described with reference to Fig. 3.

[0038] Step S100: The control unit 11 of the information processing device 10a uses the first data, the second data, and the third data as training data to train the first neural network model, the second neural network model, and the third neural network model, respectively, so that the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient satisfy the first constraint. That is, the control unit 11 trains the first neural network model, the second neural network model, and the third neural network model, respectively, under constraint conditions. The first constraint condition is a condition for training the first neural network model to represent the third data 131 based on the first weighting coefficient and the second weighting coefficient. Here, the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient include the weight itself and a bias and refer to the entire parameters of the neural network.

[0039] The first data, second data, and third data may be any data such as text, image, audio, etc. In the following description of the present embodiment, the first data, second data, and third data are all image data, but the present invention is not limited to this.

[0040] The first constraint condition may be, for example, a condition that the third weighting factor is trained to be the average value of the first weighting factor and the second weighting factor. Alternatively, the first constraint condition may be a condition that the third weighting factor is trained to be the weighted average value of the first weighting factor and the second weighting factor. Alternatively, the first constraint condition may be a condition that the third weighting factor is trained to be a value obtained by applying a function to the first and second weighting factors. The first constraint condition may be any condition that trains the third weighting factor to be represented by the first weighting factor and the second weighting factor.

[0041] 4 shows a schematic configuration of a mixed model 100 according to this embodiment. As shown in FIG. 4, the mixed model 100 according to this embodiment includes a first neural network model 110, a second neural network model 120, and a third neural network model 130. The first neural network model 110, the second neural network model 120, and the third neural network model 130 calculate the weighting coefficients θ using first data 111, second data 121, and third data 131 as training data, respectively. １ , θ2, and θ ３ is learned so as to satisfy the first constraint. １ , θ2, and θ ３ satisfies the first constraint if θ ３ is θ １ , θ2. For example, if the third weighting factor is the average value of the first and second weighting factors, then θ ３ is expressed by the following equation (1).

[0042]

[0043] Furthermore, for example, when the third weighting factor is a weighted average value of the first weighting factor and the second weighting factor, θ ３ is expressed by the following equation (2).

[0044] Here, e1 and e2 are weighting coefficients relating to the weighted average of the first weighting coefficient and the second weighting coefficient, respectively.

[0045] Specifically, the learning process involves inputting the coordinates (x, y) of each pixel into the neural network model, repeating the learning a predetermined number of times (epochs), and adjusting the weighting coefficient θ so that the loss function is minimized. １ , θ2, and θ3 are repeatedly updated. FIG. 5 shows a conceptual diagram of the learning process. The first neural network model 110 has a loss function value determined based on the error between the output obtained by inputting the coordinates (x, y) of each pixel and the first data 111 serving as training data. The second neural network model 120 has a loss function value determined based on the error between the output obtained by inputting the coordinates (x, y) of each pixel and the second data 121 serving as training data. The third neural network model 130 has a loss function value determined based on the error between the output obtained by inputting the coordinates (x, y) of each pixel and the third data 131 serving as training data. The weighting coefficient θ is adjusted so that the value calculated from these three loss functions is minimized. １ , θ2, and θ3 are updated. The value calculated from the values ​​of the three loss functions is, for example, the average value of the three loss functions, but is not limited to this. Here, the initial values ​​of the first weighting coefficient and the second weighting coefficient (θ １ , θ2) are randomly determined. Note that the initial value of the first weighting coefficient and the initial value of the second weighting coefficient are often the same, but this is not limited to this. On the other hand, the initial value of the third weighting coefficient (initial value of θ3) is determined so as to satisfy the first constraint. In this learning process, the weighting coefficients θ １, θ2 is updated based on the backpropagation method. The weighting coefficient θ3 is updated so as to satisfy the first constraint.

[0046] As a result of the learning process, the first weighting coefficient θ １ and the second weighting coefficient θ2 is updated, and the third weighting coefficient is the updated first weighting coefficient θ １ and the second weighting coefficient θ2. When the coordinates (x, y) of each pixel are input to each neural network model constituting the mixture model 100, data corresponding to the first data 111, the second data 121, and the third data 131 are output from the first neural network model 110, the second neural network model 120, and the third neural network model 130, respectively. Note that the data corresponding to the first data 111, the second data 121, and the third data 131 are data that reproduce the first data 111, the second data 121, and the third data 131, which are the teacher data. In other words, the data corresponding to the first data 111, the second data 121, and the third data 131 are data that are the same as or similar to the first data 111, the second data 121, and the third data 131, which are the teacher data.

[0047] 6 is a schematic diagram showing a mixed model in which the first constraint is that the third weighting factor is the average value of the first and second weighting factors. In FIG. 6, the first constraint is a function F α In the mixture model shown in FIG. 6, the third weighting coefficient θ ３ is expressed by the following equation (3).

[0048] As described above, the information processing device 10a can encode the first data, the second data, and the third data into the first weighting coefficient and the second weighting coefficient.

[0049] Returning to Fig. 3, the operation of the information processing device 10a will be described. Step S110: The output unit 14 of the information processing device 10a outputs the first weighting coefficient and the second weighting coefficient generated in step S100.

[0050] Next, the operation of the information processing device 10b will be described with reference to Fig. 7. The information processing device 10b decodes the data encoded by the information processing device 10a. Step S200: The input unit 13 of the information processing device 10b acquires the first weighting coefficient and the second weighting coefficient generated by the information processing device 10a.

[0051] Step S210: Next, the control unit 11 generates a third weighting factor based on the first weighting factor and the second weighting factor. The third weighting factor generated by the information processing device 10b is generated in the same manner as the third weighting factor generated by the information processing device 10a. For example, if the information processing device 10a calculates the third weighting factor by averaging the first weighting factor and the second weighting factor, the information processing device 10b also calculates the third weighting factor by averaging the first weighting factor and the second weighting factor. Step S220: Next, the control unit 11 sets the first weighting factor, the second weighting factor, and the third weighting factor to the neural network model, inputs the coordinates (x, y) of each pixel, and the obtained output becomes data corresponding to the first data, the second data, and the third data, respectively, and decodes them.

[0052] As described above, the information processing device 10 according to this embodiment trains the first neural network model, the second neural network model, and the third neural network model so that the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient satisfy constraints. With this configuration, the third data is represented based on the first weighting coefficient and the second weighting coefficient. Therefore, for example, the third data can be decoded by transmitting only the first weighting coefficient and the second weighting coefficient to the decoding-side information processing device, which is advantageous in terms of improving data transmission efficiency, etc. In other words, technology related to data representation using neural network models is improved in that the third data can be decoded by transmitting the weighting coefficients of the first data and the weighting coefficients of the second data.

[0053] In this embodiment, the mixed model includes three neural network models, but the mixed model may include four or more neural network models. FIG. 8 is a schematic diagram showing an example of a configuration in which the mixed model includes four neural network models. The mixed model 100B shown in FIG. 8 includes a first neural network model 110, a second neural network model 120, a third neural network model 130, and a fourth neural network model 140. The first neural network model 110, the second neural network model 120, the third neural network model 130, and the fourth neural network model 140 calculate weighting coefficients θ using first data 111, second data 121, third data 131, and fourth data 141 as training data, respectively. １ , θ2, and θ B is learned so as to satisfy the first constraint condition. As described above, the first constraint condition is a condition for learning so as to represent the third data 131 based on the first weighting coefficient and the second weighting coefficient. １ , θ2, and θ C is learned to satisfy the second constraint. The second constraint is a condition for learning to represent the fourth data 141 based on the first weighting coefficient and the second weighting coefficient. In FIG. 8, the first constraint is the function F β Each weighting coefficient θ １ , θ2, and θ B satisfies the first constraint if θ B is θ １ , θ2. The weighting coefficient θ of the mixture model 100B shown in FIG. B is expressed by the following equation (4).

[0054] In addition, in Figure 8, the second constraint is the function F γ Each weighting coefficient θ １ , θ2, and θ C satisfies the second constraint if θ C is θ １, θ2. The weighting coefficient θ of the mixture model 100B shown in FIG. C is expressed by the following equation (5).

[0055] As described above, in this embodiment, the first neural network model, the second neural network model, the third neural network model, and the fourth neural network model may be trained so that each weighting coefficient satisfies a constraint. According to this configuration, the third data and the fourth data are represented based on the first weighting coefficient and the second weighting coefficient. Therefore, for example, the third data and the fourth data can be decoded by transmitting only the first weighting coefficient and the second weighting coefficient to the decoding-side information processing device, which is advantageous in terms of improving data transmission efficiency, etc. In other words, technology related to data representation using neural network models is improved in that the third data and the fourth data can be decoded by transmitting the weighting coefficients of the first data and the weighting coefficients of the second data.

[0056] 9 is a schematic diagram showing a case in which, in the mixture model 100B, each training data is a partial image obtained by dividing the entire image 200 into four parts. In such a case, for example, the four partial images (partial image 211, partial image 221, partial image 231, and partial image 241) are represented based on the first weighting coefficient and the second weighting coefficient. Therefore, for example, by transmitting only the first weighting coefficient and the second weighting coefficient to the information processing device on the decoding side, all partial images can be decoded, which has the advantage of making it possible to decode the entire image 200. Note that the same processing as described above can also be performed when dividing an image into parts other than four.

[0057] In this embodiment, an example has been described in which the number of neural network models included in the mixed model is three or four, but this is not limiting. The number of neural network models included in the mixed model may be five or more. The inventors have succeeded in encoding and decoding 128 images using 128 neural network models. By using the technology disclosed herein to establish a relationship between the weighting coefficients, it is generally possible to encode and decode different weighting coefficients θ1, ..., θ ｎBased on this, m pieces of data (where m>n) can be represented. Here, the number of layers of each neural network model included in the mixed model may or may not be the same. If the number of layers of each neural network model is the same, there is an advantage in that the constraints between the weight coefficients are simplified. For example, if the mixed model includes three neural network models, the number of layers of the first neural network model, the number of layers of the second neural network model, and the number of layers of the third neural network model need only be the same. This makes it easy to set the constraints on the weight coefficients of each neural network model. On the other hand, even if the number of layers of each neural network model is not the same, it is sufficient to define a constraint equation such that θ3 can be calculated from θ1 and θ2.

[0058] When encoding image data, it is desirable that the activation function of each neural network model included in the mixture model is a sine function. By adopting a sine function as the activation function, training data can be reproduced with high accuracy. The activation function may be changed as appropriate depending on the object to be encoded. For example, the activation function may be any function, such as a ReLU function or a sigmoid function.

[0059] While the present disclosure has been described based on various drawings and examples, it should be noted that those skilled in the art may make various modifications and alterations based on the present disclosure. Therefore, it should be noted that these modifications and alterations are within the scope of the present disclosure. For example, functions included in each component or step can be rearranged so as not to cause logical inconsistencies, and multiple components or steps can be combined or divided into one. Furthermore, while the present disclosure has been described based on still images, any data that can be encoded using a neural network model, not limited to image data, can be encoded and decoded based on the present disclosure. Furthermore, the technology of the present disclosure can also be applied to image classification using general neural networks. For example, it is possible to improve adversarial attack resistance by using a third neural network corresponding to the first and second neural networks.

[0060] The technology of the present disclosure can be used in data representation technology for images and the like, data encoding technology, as well as security technology and the like.

[0061] REFERENCE SIGNS LIST 1 System 10, 10a, 10b Information processing device 11 Control unit 12 Storage unit 13 Input unit 14 Output unit 15 Communication unit 20 Network 30a, 30b Neural network model 100, 100B Mixture model 110 First neural network model 120 Second neural network model 130 Third neural network model 140 Fourth neural network model 111 First data 121 Second data 131 Third data 141 Fourth data 200 Whole image 211, 221, 231, 241 Partial image data

Claims

1. A data processing method executed by an information processing apparatus, the method comprising: using first data, second data, and third data as teacher data to train a first neural network model, a second neural network model, and a third neural network model, respectively, such that first weight coefficients related to the first neural network model, second weight coefficients related to the second neural network model, and third weight coefficients related to the third neural network model satisfy a first constraint condition; wherein the first constraint condition is a condition for training to represent the third data based on the first weight coefficients and the second weight coefficients.

2. The data processing method according to claim 1, wherein the third weight coefficient is an average value of the first weight coefficient and the second weight coefficient.

3. The data processing method according to claim 1, wherein the third weight coefficient is a weighted average value of the first weight coefficient and the second weight coefficient.

4. The data processing method according to claim 1, further comprising: using fourth data as teacher data to train a fourth neural network model such that the first weight coefficients, the second weight coefficients, and fourth weight coefficients related to the fourth neural network model satisfy a second constraint condition; wherein the second constraint condition is a condition for training to represent the fourth data based on the first weight coefficients and the second weight coefficients.

5. The data processing method according to claim 1, wherein activation functions of the first neural network model, the second neural network model, and the third neural network model are Sine functions.

6. The data processing method according to claim 1, wherein in the training step, an initial value of the first weight coefficient is the same as an initial value of the second weight coefficient.

7. The data processing method according to claim 1, wherein the number of layers of the first neural network model, the number of layers of the second neural network model, and the number of layers of the third neural network model are the same.

8. A data processing method executed by an information processing apparatus, comprising: using first data, second data, and third data as teacher data to train a first neural network model, a second neural network model, and a third neural network model, respectively, such that first weight coefficients related to the first neural network model, second weight coefficients related to the second neural network model, and third weight coefficients related to the third neural network model satisfy a first constraint condition, where the first constraint condition is a condition for training to represent the third data based on the first weight coefficients and the second weight coefficients; obtaining the first weight coefficients and the second weight coefficients; generating the third weight coefficients based on the first weight coefficients and the second weight coefficients; and performing data processing based on the third weight coefficients.

9. An information processing apparatus including a control unit, where the control unit uses first data, second data, and third data as teacher data to train a first neural network model, a second neural network model, and a third neural network model, respectively, such that first weight coefficients related to the first neural network model, second weight coefficients related to the second neural network model, and third weight coefficients related to the third neural network model satisfy a first constraint condition, and the first constraint condition is a condition for training to represent the third data based on the first weight coefficients and the second weight coefficients.

10. A program that causes a computer to use first data, second data, and third data as teacher data to train a first neural network model, a second neural network model, and a third neural network model, respectively, such that first weight coefficients related to the first neural network model, second weight coefficients related to the second neural network model, and third weight coefficients related to the third neural network model satisfy a first constraint condition, where the first constraint condition is to train the third data to be expressed based on the first weight coefficients and the second weight coefficients.

Images