JP7498421B2 - Machine learning model training method, data generation device, and trained machine learning model - Google Patents

Description

This specification relates to a machine learning model and a technique for generating data using the machine learning model.

In recent years, the use and development of machine learning models such as neural networks and support vector machines has progressed. For example, in the technology disclosed in Patent Document 1, input data is labeled with positive or negative information indicating whether or not it has a certain attribute, and used as training data. The support vector machine is then trained by having it learn the features of the training data.

JP 2013-25398 A

However, it may be desirable to adjust the progress of training of a machine learning model in relation to the learning progress of other machine learning models. With the above technology, it may not be possible to appropriately adjust the progress of such training.

This specification discloses a technique for adjusting the learning progress of a machine learning model.

The technology disclosed in this specification has been made to solve at least some of the problems described above, and can be realized as the following application examples.

[Application Example 1] A method for training a machine learning model, comprising: a first step of inputting first input data into a first machine learning model to output first output data corresponding to the first input data, the first machine learning model being a model that performs an arithmetic process on the first input data using a plurality of first arithmetic parameters to output the first output data; and a second step of adjusting the plurality of first arithmetic parameters using the first output data and teacher data, the teacher data indicating a target value of the first output data corresponding to the first input data, wherein a computer repeats the first step and the second step multiple times to train the first machine learning model, and the second step includes a varying step of varying values used for adjusting the plurality of first arithmetic parameters using random numbers.

According to the above configuration, the second step includes a varying step of varying the values used for adjusting the plurality of first calculation parameters by using random numbers, so that the adjustment amounts of the plurality of first calculation parameters can be varied to amounts different from the adjustment amounts calculated when no random numbers are used. As a result, the progress of training of the first machine learning model can be adjusted.
[Application Example 2]
The method according to Application Example 1,
The varying step is a step of varying at least a part of values included in the plurality of teacher data by using the random numbers;
The method, wherein the second step is a step of adjusting the plurality of first calculation parameters using the first output data and the teacher data after at least a portion of the values have been varied.
[Application Example 3]
The method according to Application Example 1,
the varying step is a step of varying at least a part of values included in the first output data by using the random number;
The method, wherein the second step is a step of adjusting the plurality of first calculation parameters using the first output data after at least some of the values have been changed and the teacher data.
[Application Example 4]
The method according to Application Example 1,
The second step comprises:
Calculating an evaluation value based on a predetermined loss function using the first output data and the teacher data;
the varying step of varying the evaluation value by using the random number;
adjusting the plurality of first calculation parameters using the evaluation value after the change;
A method comprising:
[Application Example 5]
The method according to any one of Application Examples 1 to 4, further comprising:
a third step of inputting second input data into a second machine learning model different from the first machine learning model, thereby outputting second output data corresponding to the second input data, the second machine learning model being a model that performs a calculation process using a plurality of second calculation parameters on the second input data to output the second output data;
a fourth step of adjusting the plurality of second calculation parameters using the second output data;
Equipped with
training the first machine learning model and the second machine learning model in parallel by repeating the first step, the second step, the third step, and the fourth step a plurality of times;
The method, wherein the fourth step does not include a step of varying values used to adjust the plurality of second operational parameters using random numbers.
[Application Example 6]
The method according to application example 5,
The first input data is a plurality of data including real data and the second output data as false data,
the teacher data is data indicating whether each of the plurality of first input data is the real data or the fake data,
the first machine learning model is a model for identifying whether each of the plurality of first input data is the real data or the fake data;
The first output data is data indicating a classification result by the first machine learning model,
the second machine learning model is a model for generating the second output data as the fake data,
In the second step, the plurality of first calculation parameters are adjusted using the first output data and the teacher data so that a difference between the first output data and the teacher data becomes small;
In the fourth step, the plurality of second calculation parameters are adjusted so that a difference between the first output data obtained by inputting the second output data as the fake data into the first machine learning model and the teacher data is increased.
[Application Example 7]
The method according to any one of Application Examples 1 to 6,
The method includes varying values used for adjusting the plurality of first calculation parameters so that the values after the variation are equal to or greater than 0 in the varying step.

[Application Example 2] A data generating device includes a generation unit that generates specific data using a learned second machine learning model trained by a training process using a first machine learning model, and an output unit that outputs the specific data, wherein the first machine learning model is a model that performs a calculation process using a plurality of first calculation parameters on first input data, and outputs identification data indicating a result of identifying whether the first input data is real data or fake data, and the second machine learning model is a model that performs a calculation process using a plurality of second calculation parameters on second input data, and generates the fake data corresponding to the second input data, and the training process includes a fake data generation process that inputs the second input data to the second machine learning model to generate the fake data, and a fake data output process that outputs identification data indicating a result of identifying whether the first input data is real data or fake data generated by the second machine learning model. a first adjustment process for adjusting the plurality of first calculation parameters using the identification data and teacher data so as to reduce a difference between the identification data and the teacher data, the teacher data being data indicating whether each of the plurality of first input data is the real data or the fake data; and a second adjustment process for adjusting the plurality of second calculation parameters using the identification data and the teacher data obtained by inputting the fake data into the first machine learning model so as to increase a difference between the identification data and the teacher data, the fake data generation process, the identification data generation process, the first adjustment process, and the second adjustment process being repeated a plurality of times to train the first machine learning model and the second machine learning model in parallel,
A data generating device, wherein the first adjustment process includes a process of varying values used to adjust the plurality of first calculation parameters using random numbers, and the second adjustment process does not include a process of varying values used to adjust the plurality of second calculation parameters using random numbers.

According to the above configuration, the first adjustment process includes a process of varying the values used to adjust the multiple first calculation parameters using random numbers, and the second adjustment process does not include a process of varying the values used to adjust the multiple second calculation parameters using random numbers. As a result, by adjusting the progress of the training of the first machine learning model, the second machine learning model is trained so as not to lag excessively behind the progress of the training of the first machine learning model. To this end, for example, the data generation device can generate appropriate output data.

The technology disclosed in this specification can be realized in various forms, such as a training device that executes the above-mentioned training method, a trained machine learning model, a computer program for implementing these methods and devices, a recording medium on which the computer program is recorded, etc.

FIG. 2 is a block diagram showing a configuration of a data generating device 200. 13 is a flowchart of an image generating process. FIG. 2 is a diagram showing an example of an input image and an output image. FIG. 2 is a block diagram showing the configuration of a generation network GN. FIG. 2 is a block diagram showing a configuration of an encoder EC. FIG. 2 is a block diagram showing the configuration of a decoder DC. FIG. 1 is a block diagram showing a configuration of a training device 100. FIG. 2 is a diagram showing an example of an input image and an actual image. FIG. 1 is a conceptual diagram of a network system 1000 according to an embodiment of the present invention. FIG. 2 is a block diagram showing the configuration of an identification network DN. 13 is a flowchart of a training process. 13 is a flowchart of a fluctuation error value calculation process. 13 is a flowchart of a fluctuation error value calculation process according to a first modified example. 13 is a flowchart of a fluctuation error value calculation process according to a second modified example.

A. Example A-1 Configuration of Data Generating Device Next, an embodiment will be described based on an example. Fig. 1 is a block diagram showing the configuration of a data generating device 200 of this example.

The data generating device 200 is a computer such as a personal computer or a smartphone. The data generating device 200 includes a CPU 210 as a controller of the data generating device 200, a volatile storage device 220 such as a RAM, a non-volatile storage device 230 such as a hard disk drive or a flash memory, a display unit 240 such as a liquid crystal display, an operation unit 250 such as a keyboard or a mouse, and a communication interface (IF) 270. The communication interface 270 is an interface for connecting to an external device (e.g., the printer 300). The communication interface 270 includes, for example, a wired or wireless interface for connecting to a network NW to which the printer 300 is connected.

The volatile storage device 220 provides a buffer area for temporarily storing various intermediate data generated when the CPU 210 performs processing. The non-volatile storage device 230 stores a computer program PGg and an input data group IG including multiple input data IDs, which will be described later. The volatile storage device 220 and the non-volatile storage device 230 are internal memories of the data generating device 200.

The computer program PGg is provided, for example, in a form that is downloaded from a server operated by the manufacturer of the printer 300. Alternatively, the computer program PGg may be provided in a form that is stored on a DVD-ROM or the like. The CPU 210 executes the computer program PGg to perform the image generation process described below.

The computer program PGg includes a generator network program GNM as a module, which is a computer program that causes the CPU 210 to realize the functions of the generator network GN described below.

The printer 300 is an inkjet or electrophotographic printing device that prints images on a printing medium such as paper using ink or toner as a printing material.

A-2. Image Generation Processing Fig. 2 is a flowchart of the image generation processing. This image generation processing is started, for example, in the data generating device 200 based on a start instruction from the user.

In S10, the CPU 210 acquires one input data ID. This input data ID is, for example, one piece of data selected based on a user's specification from the input data group IG stored in the non-volatile memory device 230. The input data ID is image data indicating the input image II.

Figure 3 shows an example of an input image and an output image. In this embodiment, input image II is an image showing a specific character in a first typeface (font). Input image II includes (H x W) pixels arranged in a matrix in the vertical and horizontal directions. H is the number of pixels in the vertical direction of input image II, and W is the number of pixels in the horizontal direction of input image II. In this embodiment, H = W = 256. For example, input image IIa in Figure 3 (A) is an image showing the character "A" in the first typeface. Input image IIb in Figure 3 (B) is an image showing the character "B" in the first typeface.

In this embodiment, the input data ID is bitmap data representing an image including multiple pixels, specifically, RGB image data that represents the color of each pixel by RGB values. The RGB values are the gradation values (hereinafter also referred to as component values) of three color components, that is, the color values of the RGB color system including the R value, the G value, and the B value. The R value, the G value, and the B value are, for example, gradation values of a predetermined number of gradations (for example, 256).

In S20, the CPU 210 inputs the input data ID to the generation network GN to generate output data OD corresponding to the input data ID. The output data OD is image data showing the output image OI, and is RGB image data like the input data ID. The output image OI in this embodiment is an image showing the same character as the corresponding input image II in a second typeface different from the first typeface. For example, the output image OIa in FIG. 3(A) is an output image corresponding to the input image IIa, and is an image showing the character "A" in the second typeface. The output image OIb in FIG. 3(B) is an output image corresponding to the input image IIb, and is an image showing the character "B" in the second typeface. In this embodiment, the size of the output image OI is the same as the size of the input image II. The specific configuration of the generation network GN will be described later.

In S30, the CPU 210 outputs the generated output data OD. For example, the CPU 210 generates print data indicating the output image OI using the output data OD and transmits it to the printer 300. The printer 300 prints the output image OI on a print medium using the print data. Alternatively, the CPU 210 displays the output image OI on the display unit 240 using the output data OD.

A-3. Configuration of the generation network GN Fig. 4 is a block diagram showing the configuration of the generation network GN. The generation network GN includes an encoder EC and a decoder DC.

A-3-1. Configuration of Encoder EC The encoder EC executes a dimensionality reduction process on the input data ID using a plurality of calculation parameters Pe to generate feature data CD indicating the features of the input data ID (i.e., the features of the input image II). In this embodiment, the input data ID includes three component values (R value, G value, B value) of each of the (256×256) pixels, so that the data includes (256×256×3) values, that is, data of (256×256×3) dimensions. In this embodiment, the feature data CD includes (1×1×512) values, that is, data of 512 dimensions. In this way, the number of dimensions of the input data ID is reduced in the dimensionality reduction process.

Figure 5 is a block diagram showing the configuration of the encoder EC. The encoder EC is a neural network that has an input layer EL_0 and multiple convolution layers EL_1 to EL_8.

The input layer EL_0 is a layer to which the input data ID is input. The first convolutional layer EL_1 receives the input data ID input to the input layer EL_0 as is. The convolutional layer EL_1 performs a calculation process, described later, on the (256×256×3)-dimensional input data ID to generate (A ₁ ×B ₁ ×C ₁ )-dimensional data (A ₁ , B ₁ , and C ₁ are positive integers).

The kth (k is an integer from 2 to 8) convolutional layer EL_k receives (A _{k-1 , B k-1 , C k-1 )-dimensional processed data obtained by performing a predetermined post-processing (described later) on (A k-1 ×} B _{k-1 ×} C _k-1 )-dimensional data generated by the (k-1)th convolutional layer EL_ _(k-1 ). The convolutional layer EL_k performs arithmetic processing (described later) on the (A _k-1 ×B _k-1 ×C _k-1 )-dimensional processed data to generate (A _k ×B _k ×C _k )-dimensional data (A _k , B _k , C _k are positive integers).

The computational processes performed by each of the convolutional layers EL_1 to EL_8 include convolutional processing and bias addition processing. The convolutional processing is a process of sequentially applying s filters of (p×q×r) dimensions to input data to calculate a correlation value indicating the correlation between the input data and the filters. In the process of applying each filter, multiple correlation values are sequentially calculated while sliding the filter. One filter includes (p×q×r) weights. The bias addition processing is a process of adding biases, one for each filter, to the calculated correlation value. The (p×q×r×s) weights included in the s filters and the s biases corresponding to the s filters are the multiple computational parameters Pe described above, and are updated in the training processing described below.

Each value of the data generated by each of the convolution layers EL_1 to EL_8 is a value obtained by adding a bias to the correlation value described above. The number of data included in the data generated by each of the convolution layers EL_1 to EL_8 (for example, (A ₁ ×B ₁ ×C ₁ ) in the case of the convolution layer EL_1) is determined by the stride (the amount by which the filter is slid) in the convolution process and the number of filters s.

The data values generated by the convolutional layer EL_1 are input to an activation function and transformed as the post-processing described above. In this embodiment, the activation function used is the so-called LeakyReLU (Leaky Rectified Linear Unit).

The data values generated by convolutional layers EL_2 to EL_8 are transformed by batch normalization as the post-processing mentioned above, and then input to an activation function for further transformation. Batch normalization is a process in which the average and variance of each value are calculated for a set (batch) of input data used in the training process described below, and each value is normalized. In the image generation process, each value is normalized using the moving average value of the average and variance calculated for each batch during the training process.

The processed data obtained by performing the above-mentioned post-processing on the data generated by the convolutional layer EL_8 is the above-mentioned feature data CD.

In this embodiment, the numbers of dimensions (A ₁ ×B ₁ ×C ₁ ) to (A ₈ ×B ₈ ×C ₈ ) of the data generated by each of the convolution layers EL_1 to EL_8 are as follows:
( _A1 x _B1 x _C1 ) = (128 x 128 x 64)
( _A2 x _B2 x _C2 ) = (64 x 64 x 128)
( _A3 x _B3 x _C3 ) = (32 x 32 x 256)
( _A4 x _B4 x _C4 ) = (16 x 16 x 512)
( _A5 x _B5 x _C5 ) = (8 x 8 x 512)
( _A6 x _B6 x _C6 ) = (4 x 4 x 512)
( _A7 x _B7 x _C7 ) = (2 x 2 x 512)
( _A8 x _B8 x _C8 ) = (1 x 1 x 512)

A-3-2. Configuration of the decoder DC The decoder DC executes a dimension restoration process on the feature data CD generated by the encoder EC using a plurality of calculation parameters Pd to generate the above-mentioned output data OD. In this embodiment, the feature data CD is data including (1×1×512) values as described above, that is, 512-dimensional data. In this embodiment, the output data OD is data including (256×256×3) values, that is, (256×256×3)-dimensional data, similar to the input data ID. In this manner, in the dimension restoration process of this embodiment, the number of dimensions of the feature data CD is restored.

Figure 6 is a block diagram showing the configuration of the decoder DC. The decoder DC is a neural network having multiple transposed convolution layers DL_1 to DL_8.

Feature data CD is input to the first transposed convolutional layer DL_1. The transposed convolutional layer DL_1 performs arithmetic processing, which will be described later, on the feature data CD to generate ( _D1 x _E1 x _F1 )-dimensional data ( _D1 , _E1 , _F1 are positive integers).

The m-th (m is an integer from 2 to 8) transposed convolution layer DL_m receives processed data of (D _m-1 , E _m- 1 , F m-1 ) dimensions obtained by performing a predetermined post-processing (described later) on data of (D m _-1 , E _m-1 , F m-1 ) dimensions generated by the (m-1)th transposed convolution layer DL_( _{m -1} ). Furthermore, the m-th transposed convolution layer DL_m receives processed data of (A 9- _m , B _{9-m , C 9-m ) dimensions obtained by performing the above-mentioned post-processing on data generated by the (9-m} )th convolution layer EL_ _(9-m ) of the encoder EC. For example, as shown in FIG. 5 and FIG. 6, the transposed convolution layer DL_2 of the decoder DC receives processed data based on data generated by EL_7 of the encoder EC, and the transposed convolution layer DL_4 receives processed data based on data generated by EL_5 of the encoder EC. Therefore, {(D _m-1 , E _m-1 , F _m-1 )+(A _9-m , B _9-m , C _9-m )}-dimensional processed data is input to the m-th (m is an integer from 2 to 8) transposed convolution layer DL_m. The transposed convolution layer DL_m performs arithmetic processing, described later, on the input processed data to generate (D _m ×E _m ×F _m )-dimensional data (D _m , E _m , F _m are positive integers).

The computational processes performed by each of the transposed convolution layers DL_1 to DL_8 include transposed convolution and bias addition. The transposed convolution process is a process in which a value (e.g., a value of zero) is added to the input data as appropriate according to the stride to increase the number of dimensions, and then a convolution operation is performed using a (p×q×r)-dimensional filter, similar to the above-mentioned convolution process. The bias addition process is a process in which a bias prepared for each filter is added to the correlation value calculated by the transposed convolution operation. The (p×q×r×s) weights included in the s filters and the s biases corresponding to the s filters are the above-mentioned multiple computational parameters Pd, and are updated in the training process described below.

Each value of the data generated by each of the transposed convolution layers DL_1 to DL_8 is a value obtained by adding a bias to the correlation value described above. The number of data included in the data generated by each of the transposed convolution layers DL_1 to DL_8 (for example, (D ₁ ×E ₁ ×F ₁ ) in the case of the transposed convolution layer DL_1) is determined by the stride (the amount of values such as zeros added) in the transposed convolution process and the number of filters s.

The values of the data generated by the transposed convolution layers DL_1 to DL_3 are transformed by the batch normalization described above as the post-processing described above. Then, in the training process, each value transformed by batch normalization is further transformed by dropout as post-processing, and then input to the activation function and transformed. Dropout is a process that randomly invalidates (sets to 0) some values to suppress overlearning. The activation function used is the so-called ReLU (Rectified Linear Unit). In the image generation process, dropout processing is not performed, and each value transformed by batch normalization is input to the activation function and transformed.

The data values generated by the transposed convolution layers DL_4 to DL_7 are transformed by the batch normalization described above as the post-processing described above, and then input to the activation function and transformed. In the post-processing of the transposed convolution layers DL_4 to DL_7, dropout is not performed in either the training process or the image generation process.

The ( _D8 × _E8 × _F8 )-dimensional data generated by the transposed convolution layer DL_8 is the above-mentioned output data OD. Therefore, the number of dimensions ( _D8 × _E8 × _F8 ) of the data generated by the transposed convolution layer DL_8 is equal to the number of dimensions (256×256×3) of the output data OD.

In this embodiment, the numbers of dimensions (D ₁ ×E ₁ ×F ₁ ) to (D ₈ ×E ₈ ×F ₈ ) of the data generated by each of the transposed convolution layers DL_1 to DL_8 are as follows:
( _D1 x _E1 x _F1 ) = (2 x 2 x 512)
( _D2 x _E2 x _F2 ) = (4 x 4 x 512)
( _D3 x _E3 x _F3 ) = (8 x 8 x 512)
( _D4 x _E4 x _F4 ) = (16 x 16 x 512)
( _D5 x _E5 x _F5 ) = (32 x 32 x 256)
( _D6 x _E6 x _F6 ) = (64 x 64 x 128)
( _D7 x _E7 x _F7 ) = (128 x 128 x 64)
( _D8 x _E8 x _F8 ) = (256 x 256 x 3)

A-4. Training of the Generative Network GN The above-mentioned generative network GN is trained so that it can generate the desired output data OD when the input data ID is input. The training of the generative network GN will be described below.

A-4-1. Configuration of the Training Apparatus Fig. 7 is a block diagram showing the configuration of a training apparatus 100 that executes training of the generative network GN of this embodiment.

The training device 100 is a computer such as a personal computer or a server. The training device 100 includes a CPU 110 as a controller for the training device 100, a volatile storage device 120 such as a RAM, a non-volatile storage device 130 such as a hard disk drive or a flash memory, a display unit 140 such as a liquid crystal display, an operation unit 150 such as a keyboard or a mouse, and a communication interface (IF) 170 for connecting to an external device (e.g., a printer 300).

The volatile memory device 120 provides a buffer area for temporarily storing various intermediate data generated when the CPU 110 performs processing. The non-volatile memory device 130 stores a computer program PGt, an input data group IG including multiple input data ID for learning, an actual data group RG including multiple actual data RD, and teacher data LD. The volatile memory device 120 and the non-volatile memory device 130 are internal memories of the training device 100.

The computer program PGt is provided, for example, in a form that is downloaded from a server operated by the manufacturer of the printer 300. Alternatively, the computer program PGt may be provided in a form that is stored on a DVD-ROM or the like. The CPU 110 executes the computer program PGt to perform the training process described below.

The input data ID used in the training process is RGB image data similar to the input data ID used in the image generation process described above. The input image II indicated by the input data ID is an image showing a specific character in a first typeface (font), as described above. The actual data RD is RGB image data similar to the output data OD generated in the image generation process described above. The multiple actual data RD correspond one-to-one to each of the multiple input data IDs included in the input data group IG. In this embodiment, the actual image RI indicated by the actual data RD is an image showing the same character as the corresponding input image II in a second typeface.

Figure 8 shows an example of an input image and an actual image. Input image IIc in Figure 8 (A) is an image showing the letter "C" in a first typeface. Actual image RIc corresponding to input image IIc in Figure 8 (A) is an image showing the letter "C" in a second typeface. Input image IId in Figure 8 (B) is an image showing the letter "d" in a first typeface. Actual image RId corresponding to input image IId in Figure 8 (B) is an image showing the letter "d" in the second typeface.

A predetermined number of pairs of input data ID and real data RD, for example 1,000, are prepared by the worker performing the training.

The teacher data LD is data that indicates the target value of the identification data DD that should be output by the identification network DN, which will be described later. The teacher data LD is prepared by the operator who performs the training. The teacher data LD will be described further below.

A-4-2. Configuration of the network system FIG. 9 is a conceptual diagram of a network system 1000 of this embodiment. The network system 1000 is a system used to train the generative network GN, and includes a discriminator network DN in addition to the generative network GN described above. The generative network GN and the discriminator network DN constitute a so-called generative adversarial network (GANs (generative adversarial networks)).

A pair of data (also called a real data pair Pr) consisting of an input data ID and real data corresponding to the input data ID is input to the identification network DN. Furthermore, a pair of data (also called a fake data pair Pf) consisting of an input data ID and an output data OD corresponding to the input data ID is input to the identification network DN. The output data OD corresponding to the input data ID means the output data OD generated by inputting the input data to the generative network GN. In addition, in a generative adversarial network, the output data OD is also called "fake data OD". Here, the fake data OD generated by inputting the input data ID corresponding to a specific real data RD to the generative network GN is also called the fake data OD corresponding to the specific real data RD.

When either a real data pair Pr or a false data pair Pf is input, the identification network DN identifies whether the input data pair is true or false. That is, the identification network DN performs an arithmetic process using a plurality of arithmetic parameters Pdn on the input data pair, and outputs identification data DD indicating the result of identifying whether the input data pair is a real data pair or a false data pair.

In this embodiment, the identification data DD is (30×30×1) dimensional data including (30×30×1) values. Hereinafter, the identification data DD output when a real data pair Pr is input is also referred to as the identification data DDr corresponding to the real data pair Pr, and the identification data DD output when a false data pair Pf is input is also referred to as the identification data DDf corresponding to the false data pair Pf. The identification network DN is trained so that the identification data DDr corresponding to the real data pair Pr indicates that the input data pair is the real data pair Pr, and the identification data DDf corresponding to the false data pair Pf indicates that the input data pair is the false data pair Pf.

For this reason, in this embodiment, two types of teacher data LD (Figure 7) are prepared: teacher data LDr corresponding to the real data pair Pr, and teacher data LDf corresponding to the false data pair Pf. Teacher data LDr is data indicating that the input data pair is the real data pair Pr, and is (30 x 30 x 1) dimensional data in which all values are "1". Teacher data LDf is data indicating that the input data pair is the false data pair Pf, and is (30 x 30 x 1) dimensional data in which all values are "0".

If the identification data DD is one-dimensional data, in the process of reducing the number of dimensions in the identification network DN to one dimension, information indicating global characteristics of the image represented by the input real data or false data may be lost. In this case, training of the identification network DN may not progress. Furthermore, if the identification data DD is one-dimensional data, the number of layers of the identification network DN, and therefore the number of calculation parameters, becomes excessively large, and the time required for training becomes excessively long. In this embodiment, by making the identification data DD data of (30 x 30 x 1) dimensions, it is possible to suppress the loss of information for identifying the authenticity of the input data pair and realize rapid training.

Figure 10 is a block diagram showing the configuration of the classification network DN. The classification network DN is a neural network that has an input layer L_0 and multiple convolution layers L_1 to L_5.

The input layer L_0 is a layer to which either the false data pair Pf or the real data pair Pr is input. The data pair input to the input layer L_0 is input as is to the first convolutional layer L_1. The number of dimensions of the input data ID constituting the data pair is (256×256×3), and the number of dimensions of the real data RD or false data OD constituting the data pair is (256×256×3). Therefore, the number of dimensions of the data pair is (256×256×6). The convolutional layer L_1 performs a calculation process described later on the (256×256×6)-dimensional input data ID to generate (G ₁ ×H ₁ ×I ₁ )-dimensional data (G ₁ , H ₁ , I ₁ are positive integers).

The nth (n is an integer from 2 to 5) convolutional layer L_n receives (G _n-1 , H n-1 , I n-1 )-dimensional processed data obtained by performing a predetermined post-processing (described later) on (G n- ₁ ×H _{n-1 ×} I _n-1 )-dimensional data generated by the (n-1)th convolutional layer L_ _(n-1 ). The convolutional layer L_n performs arithmetic processing (described later) on the (G _n-1 ×H _n-1 ×I _n-1 )-dimensional processed data to generate (G _n ×H _n ×I _n )-dimensional data (G _n , H _n , I _n are positive integers).

The computational processing performed by each of the convolutional layers L_1 to L_5 includes convolutional processing and bias addition processing, similar to the computational processing of the convolutional layer of the encoder EC. The multiple weights included in the filters used in the convolutional processing and the biases corresponding to each filter are the multiple computational parameters Pe described above, and are updated in the training processing described below. In the convolutional processing, zero padding is appropriately performed to supplement the data in order to adjust the number of dimensions of the generated data.

The data values generated by the convolutional layer L_1 are input to an activation function and transformed as the post-processing described above. In this embodiment, the activation function used is the so-called LeakyReLU (Leaky Rectified Linear Unit).

The data values generated by convolutional layers L_2 to L_4 are transformed by batch normalization as the post-processing described above, and then input to an activation function for further transformation. Batch normalization is as described above in the description of decoder DC.

The ( _G5 × _H5 × _I5 )-dimensional data generated by the convolutional layer L_5 is the above-mentioned identification data DD. Therefore, the number of dimensions ( _G5 × _H5 × _I5 ) of the data generated by the convolutional layer L_5 is equal to the number of dimensions (30×30×1) of the identification data DD.

In this embodiment, the dimensionality (G ₁ ×H ₁ ×I ₁ ) to (G ₅ ×H ₅ ×I ₅ ) of the data generated by each of the convolution layers L_1 to L_8 is as follows:
( _G1 x _H1 x _I1 ) = (128 x 128 x 64)
( _G2 x _H2 x _I2 ) = (64 x 64 x 128)
( _G3 x _H3 x _I3 ) = (32 x 32 x 256)
( _G4 x _H4 x _I4 ) = (31 x 31 x 512)
( _G5 x _H5 x _I5 ) = (30 x 30 x 1)

A-4-3. Training process The training process will be described below. The training process is a process for training the generation network GN so that it can output appropriate output data (fake data) OD by adjusting the above-mentioned multiple calculation parameters Pe and Pd of the generation network GN and the above-mentioned multiple calculation parameters Pdn of the identification network DN. The generation network GN incorporated in the computer program PGg stored in the non-volatile storage device 230 of the above-mentioned data generation device 200 is a learned model trained by this training process.

Figure 11 is a flowchart of the training process. Figure 11 is a flowchart of the training process. In S200, multiple calculation parameters Pe, Pd of the generation network GN and multiple calculation parameters Pdn of the identification network DN are initialized. For example, the initial values of these calculation parameters Pe, Pd, Pdn are set to random numbers independently obtained from the same distribution (e.g., normal distribution).

In S205, the CPU 210 selects a batch size of actual data pairs Pr, for example, V (V is an integer equal to or greater than 2, for example, V=100) from the multiple (e.g., 1000) actual data pairs Pr stored in the non-volatile storage device 230. The multiple actual data pairs Pr stored in the non-volatile storage device 230 are divided in advance into multiple groups (batches) each including V actual data pairs Pr. The CPU 210 selects V actual data pairs Pr to be used by sequentially selecting one group from the multiple groups. Alternatively, V actual data pairs Pr may be randomly selected each time from the multiple actual data pairs Pr stored in the non-volatile storage device 230.

In S210, the CPU 210 inputs each of the V input data IDs contained in the selected V real data pairs Pr to the generation network GN to generate V fake data ODs.

In S215, the CPU 210 inputs V actual data pairs Pr to the identification network DN and generates V identification data DDr corresponding to the V actual data pairs Pr.

In S220, the CPU 210 inputs V false data pairs Pf to the identification network DN to generate V identification data DDf corresponding to the V false data pairs Pf. Each of the V false data pairs Pf is a pair of V input data IDs included in the V real data pairs Pr and false data OD corresponding to the input data ID. The false data OD corresponding to the input data ID means the false data OD generated by inputting the input data ID to the generation network GN in S210.

In S225, the CPU 210 executes a variation error value calculation process. The variation error value calculation process is a process for calculating a variation error value using random numbers in addition to the teacher data LDr, LDf and the identification data DDr, DDf.

Figure 12 is a flowchart of the variation error value calculation process. In S300, the CPU 210 selects one target identification data from the V identification data DDr corresponding to the V actual data pairs Pr generated in S215 of Figure 11.

In S305, random numbers Rr are obtained for the number of dimensions of the teacher data LDr corresponding to the real data pair Pr. In this embodiment, the teacher data LDr is (30 x 30 x 1) dimensional data. Therefore, (30 x 30 x 1) random numbers Rr are obtained. These random numbers Rr are obtained independently from the same distribution, for example. In this embodiment, these random numbers Rr are obtained from a normal distribution with a mean of 0 and a variance of 0.1. Therefore, the random numbers Rr are generally values in the range of -0.3 to 0.3.

In S310, the CPU 210 generates the variation teacher data LDrv by adding the (30×30×1) random numbers Rr one by one to each of the (30×30×1) values (all "1" in this embodiment) of the teacher data LDr. The (30×30×1) values of the variation teacher data LDrv are defined as trv _i , and the (30×30×1) random numbers Rr are defined as Rr _i (i is an integer between 1 and (30×30×1)). In this case, trv _i is (1+Rr _i ).

In S315, the CPU 210 calculates the variation error value EVr using the target identification data and the variation teacher data LDrv. A predetermined loss function, in this embodiment, the sigmoid cross entropy error, is used for the variation error value EVr. The formula (1) for the sigmoid cross entropy is as follows:

Here, p is the number of dimensions of the recognition data and the teacher data (in this embodiment, p = (30 x 30 x 1)). _{a i} is a value normalized by inputting each value of the recognition data of interest into a sigmoid function, and t _i is each value of the teacher data corresponding to the value a _i . The error value Egan becomes larger as the recognition data approaches the corresponding teacher data. In other words, the error value Egan becomes larger as the difference between the recognition data and the corresponding teacher data becomes smaller. In this step, the error value Egan calculated using each value trv _i of the above-mentioned variable teacher data LDrv as each value t _i of the teacher data in equation (1) is the variable error value EVr.

In S320, the CPU 210 determines whether or not all V pieces of identification data DDr have been processed. If there is unprocessed identification data DDr (S320: NO), the CPU 210 returns to S300. If all V pieces of identification data DDr have been processed (S320: YES), the CPU 210 proceeds to S325.

When the process advances to S325, V variation error values EVr corresponding to V pieces of identification data DDr have been calculated. Each time a variation error value EVr is calculated, the random number Rr is recalculated in S305, so the variation teacher data LDrv used when calculating each of the V variation error values EVr are different from each other.

In S325, the CPU 210 selects one piece of target identification data from the V pieces of identification data DDf corresponding to the V pieces of false data pairs Pf generated in S220 of FIG. 11.

In S330, random numbers Rf are obtained for the number of dimensions of the teacher data LDf corresponding to the false data pair Pf. In this embodiment, the teacher data LDf is (30 x 30 x 1) dimensional data. Therefore, (30 x 30 x 1) random numbers Rf are obtained. These random numbers Rf are obtained independently from a normal distribution with a mean of 0 and a variance of 0.1, similar to the random number Rr described above.

In S335, the CPU 210 generates the variation teacher data LDfv by adding the absolute values |Rf| of the (30×30×1) random numbers Rf one by one to each of the (30×30×1) values (all "0" in this embodiment) of the teacher data LDf. The (30×30×1) values of the variation teacher data LDfv are designated as tfv _i , and the (30×30×1) random numbers Rf are designated as Rf _i (i is an integer between 1 and (30×30×1)). In this case, tfv _i is |Rf _i |.

In S340, the CPU 210 calculates the variation error value EVf using the target identification data and the variation teacher data LDfv. The variation error value EVf uses a sigmoid cross entropy error. Therefore, the error value Egan calculated using each value tfv _i of the variation teacher data LDfv described above as each value t _i of the teacher data in the above formula (1) is the variation error value EVf.

In S345, the CPU 210 determines whether or not all V pieces of identification data DDf have been processed. If there is unprocessed identification data DDf (S345: NO), the CPU 210 returns to S325. If all V pieces of identification data DDf have been processed (S345: YES), the CPU 210 ends the variation error value calculation process.

When the variation error value calculation process is completed, V variation error values EVf corresponding to V pieces of identification data DDf have been calculated. Each time a variation error value EVf is calculated, the random number Rf is retaken in S330, so the variation teacher data LDfv used when calculating each of the V variation error values EVf are different from each other.

When the variation error value calculation process is completed, in S230 of FIG. 11, the CPU 210 adjusts the multiple calculation parameters Pdn of the identification network DN using the calculated V variation error values EVr and V variation error values EVf. Specifically, a loss function Ldn indicating the average value of the (2×V) variation error values EVr and EVf is used as the loss function for the entire batch. The CPU 210 adjusts the calculation parameters Pdn according to a predetermined algorithm so that the index value calculated using the loss function Ldn becomes large. For example, an algorithm using backpropagation and gradient descent is used as the predetermined algorithm.

In S235, the CPU 210 calculates an error value Ef between the identification data DDf and the teacher data LDf for each of the V false data pairs Pf. Each value of the teacher data LDf is "0" as described above. Therefore, for each of the V identification data DDf, the error value Egan calculated using "0" as each value t _i of the teacher data in the above-mentioned formula (1) is the error value Ef.

In S240, the CPU 210 calculates an error value Erf between the real data RD and the fake data OD corresponding to the real data RD for each of the V real data RD. A predetermined loss function, in this embodiment, the mean absolute error (MAE (Mean Absolute Error)), is used for the error value Erf. For example, the error value Erf is calculated using the following formula (2).

Here, s is the number of dimensions of the real data RD and the fake data OD (in this embodiment, s = (256 x 256 x 3)). _{d i} is each value of the real data RD, and e _i is each value of the output data OD corresponding to the value d _i . As can be seen from equation (2), the error value Erf is the error for each pixel between the image represented by the real data RD and the image represented by the fake data OD. The closer the fake data OD and the real data RD are, the smaller the error value Erf becomes. In other words, the smaller the difference between the fake data OD and the real data RD, the smaller the error value Ef becomes.

In S245, the CPU 210 adjusts the multiple calculation parameters Pe, Pd of the generative network GN using the V error values Er calculated in S235 and the V error values Erf calculated in S240. Specifically, the CPU 210 calculates an index value using the loss function Lg shown in the following formula (3) as a loss function for the entire batch. The CPU 210 adjusts the calculation parameters Pe, Pd according to a predetermined algorithm so that the index value calculated using the loss function Lg becomes small. For example, an algorithm using backpropagation and gradient descent is used as the predetermined algorithm.

Here, as described above, Lf is the average value of V error values Ef, Lrf is the average value of V error values Erf, and α is a predetermined coefficient.

In S250, the CPU 210 judges whether the training is completed. In this embodiment, if a completion instruction is input from the operator, the training is judged to be completed, and if an instruction to continue the training is input, the training is judged to be not completed. For example, the CPU 210 inputs a plurality of test input data IDs other than the input data IDs used for training to the generation network GN to generate a plurality of false data ODs. The CPU 210 displays an image represented by the generated false data OD on the display unit 240. The operator checks the displayed image to check whether an image sufficient as a final image (for example, an image that accurately expresses a specific character in the second font) has been generated, and inputs an instruction to complete or continue the training via the operation unit 250 depending on the confirmation result. In a modified example, for example, when the processes of S205 to S245 are repeated a predetermined number of times, the training may be judged to be completed.

If it is determined that the training is not complete (S250: NO), the CPU 210 returns the process to S205. If it is determined that the training is complete (S250: YES), the CPU 210 ends the training process. According to this training process, a plurality of calculation parameters Pdn of the identification network DN are adjusted so that the variation error value EVr is increased, in other words, the difference between the identification data DDr, DDf and the variation teacher data LDrv, LDfv is reduced (S230 in FIG. 11). That is, the identification network DN is trained for the purpose of correctly identifying the truth or falsehood of the input data pair. Then, a plurality of calculation parameters Pe, Pd of the generation network GN are adjusted so that the error value Ef is reduced, in other words, the difference between the identification data DDf and the teacher data LDf is increased (S230 in FIG. 11). At the same time, the generative network GN is trained to reduce the error value Erf, in other words, to reduce the difference between the real data RD and the fake data OD corresponding to the real data RD. That is, the generative network GN is trained for the purpose of causing the identification network DN to erroneously identify the fake data pair Pf as the real data pair Pr, and for the fake data OD to approach the real data RD. As a result, the generative network GN can be trained to be able to convert another attribute to be converted (e.g., font) while retaining a specific attribute of the input data ID (e.g., the type of character to be represented). At the time the training process is completed, the generative network GN becomes a trained model with the calculation parameters Pe and Pd adjusted. Therefore, it can be said that the training process is a process of generating (manufacturing) a trained model.

According to the present embodiment described above, the training process (FIG. 11) includes a first step (S215, S220 in FIG. 11) of inputting the real data pair Pr and the false data pair Pf into the identification network DN to output the identification data DDr, DDf corresponding to these data pairs, and a second step (S225, S230 in FIG. 11) of adjusting the multiple calculation parameters Pdn of the identification network DN using the identification data DDr, DDf and the teacher data LDr, LDf, and trains the identification network DN by repeating the first and second steps multiple times (S250 in FIG. 11). The second step of the variation error value calculation process (FIG. 12) includes a variation step (S305, S310, S330, S335 in FIG. 12) of varying the teacher data LDr, LDf, which are values used to adjust the multiple calculation parameters Pdn, using random numbers Rr, Rf. As a result, the variation error value EVr varies by an amount different from the error value calculated when the random numbers Rr and Rf are not used. Therefore, the adjustment amount of the multiple calculation parameters Pdn can be varied by an amount different from the adjustment amount calculated when the random numbers Rr and Rf are not used. As a result, the progress of the training of the identification network DN can be adjusted.

For example, in this embodiment, the method further includes a third step (S210 in FIG. 11) of inputting the input data ID into a generating network GN, which is a machine learning model different from the identification network DN, to output fake data (output data) OD corresponding to the input data ID, and a fourth step (S235 to S245 in FIG. 11) of adjusting a plurality of calculation parameters Pe and Pd of the generating network GN using the fake data OD (in this embodiment, using the identification data DDf obtained by inputting the fake data OD into the identification network DN), and by repeating the first to fourth steps a plurality of times (S250 in FIG. 11), the identification network DN and the generating network GN are trained in parallel. In addition, random numbers are not used in the fourth step. In other words, the fourth step does not include a step of varying the values used to adjust the plurality of calculation parameters Pe and Pd of the generating network GN using random numbers. In this way, when multiple machine learning models are trained in parallel, for example, it may be necessary to adjust the progress of training of one machine learning model (in this embodiment, the discrimination network DN) in relation to the progress of training of the other machine learning model (in this embodiment, the generation network GN). In such a case, as in this embodiment, random numbers are not used for the values used for training (adjusting parameters) one machine learning model, but random numbers are used for the values used for training the other machine learning model, thereby making it possible to delay the progress of training of one machine learning model relative to the training of the other machine learning model.

More specifically, in this embodiment, the identification network DN and the generation network GN constitute so-called GANs, as described above. That is, the identification network DN is a model for identifying whether each of the multiple false data pairs Pf and real data pairs Pr contains real data RD or false data OD. The data output by the identification network DN is identification data DDr, DDf indicating the identification result. And the generation network GN is a model for generating the false data OD. In the second step (S225, S230 in FIG. 11), multiple calculation parameters Pdn of the identification network DN are adjusted so that the difference between the identification data DDr, DDf and the teacher data LDr, LDf is reduced (S230 in FIG. 11). In the fourth step (S235 to S245), a plurality of calculation parameters Pe, Pd of the generating network GN are adjusted so that the difference between the identification data DDf obtained by inputting the false data OD to the identification network DN and the teacher data LDf becomes large. In such GANs, the task of the generating network GN (e.g., generating image data) is heavier than the task of the identification network DN (e.g., identifying image data). For this reason, the progress of the training of the generating network GN tends to lag behind the progress of the training of the identification network DN. If, in the training process, the training of the generating network GN may be excessively delayed relative to the training of the identification network DN. In this case, when adjusting the plurality of calculation parameters Pe, Pd of the generating network GN, a problem may occur in which the gradient disappears and the training of the generating network GN does not progress. In this embodiment, the progress of the training of the identification network DN can be delayed by using the random numbers Rr, Rf in the second step. As a result, it is possible to prevent the progress of training of the generation network GN from being excessively delayed compared to the progress of training of the discrimination network DN. As a result, it is possible to effectively proceed with training of the generation network GN while suppressing the above-mentioned inconveniences.

Furthermore, in the above training process, the variation process (S305, S310, S330, S335 in FIG. 12) is a process in which the values contained in the multiple pieces of teacher data LDr, LDf are varied using random numbers Rr, Rf. As a result, by varying the values contained in the teacher data LDr, LDf, it is possible to easily adjust the progress of training of the identification network DN.

Furthermore, in the variation step of this embodiment, the values included in the teacher data LDr and LDf are varied so that the values of the teacher data after variation (i.e., the values of the varied teacher data LDrv and LDfv) are 0 or more. For example, in S335 of the above embodiment, since each value of the teacher data LDf is "0", if Rf, which has a value in the range of -0.3 to 0.3, is added as is, the value after variation may be less than 0. In S335 of the above embodiment, the absolute value |Rf| of the random number Rf is added to each value of the teacher data LDf, so that the value after variation becomes a value of 0 or more. Since values of the teacher data that are 0 or more (e.g., 0 or 1) are usually used, if the value after variation is less than 0, there is a possibility that a malfunction will occur in the calculation by the program. In this embodiment, such inconvenience can be suppressed.

Furthermore, according to the training device 100 (FIG. 1) and the trained generative network GN of this embodiment, the generative network GN is effectively trained by the above-mentioned training process, so that it can generate appropriate fake data OD according to the input data ID. For example, in this embodiment, the input data ID is image data that has a first attribute (e.g., an attribute of being the letter "A") and a second attribute (e.g., an attribute of being the first font), and does not have a third attribute (e.g., an attribute of being the second font). And the output data (fake data) OD is image data that has a first attribute and a third attribute, and does not have the second attribute. Thus, according to this embodiment, the trained generative network GN can generate the output data (fake data) OD by converting the second attribute of the input data ID into the third attribute while maintaining the first attribute of the input data ID.

As can be seen from the above explanation, the real data pair Pr and the fake data pair Pf in this embodiment are an example of the first input data, and the input data ID is an example of the second input data. Furthermore, the identification data DDr and DDf are an example of the first output data, and the fake data OD is an example of the second output data. Furthermore, the calculation parameters Pdn of the identification network DN are an example of the first calculation parameters, and the calculation parameters Pe and Pd of the generation network GN are an example of the second calculation parameters. The identification network DN is an example of a first machine learning model, and the generation network GN is an example of a second machine learning model.

B. Variations:
(1) The varying step in the above embodiment (S305, S310, S330, S335 in FIG. 12) is a step of varying values included in the multiple pieces of teacher data LDr, LDf using random numbers Rr, Rf. Not limited to this, the varying step may be a step of varying another value used for adjusting the multiple calculation parameters Pdn using random numbers. An example of this will be described as a first modified example and a second modified example.

Figure 13 is a flowchart of the variation error value calculation process of the first modified example. In the variation error value calculation process of Figure 13, S310B, S315B, S335B, and S340B are executed instead of S310, S315, S335, and S340 of the variation error value calculation process of Figure 12.

In S310B, the CPU 210 generates varying attention identification data by adding one random number Rr to each of the (30×30×1) values of the attention identification data (one of the V pieces of identification data DDr). The (30×30×1) values of the attention identification data are designated _Vri , and the (30×30×1) random numbers Rr are _designated Rri (i is an integer between 1 and (30×30×1)). In this case, each value _Vri of the varying attention identification data is ( _Vri + _Rri ).

In S315B, the CPU 210 calculates a fluctuation error value EVrb using the fluctuation attention identification data generated in S310B and the teacher data LDr. Specifically, the error value Egan calculated using the teacher data LDr value "1" as each value t _i of the teacher data in the above formula (1) and the fluctuation attention identification data value (Vr _i + Rr _i ) normalized by inputting it into a sigmoid function as a _i is the fluctuation error value EVrb.

In S335B, the CPU 210 generates varying attention identification data by adding one random number Rf to each of the (30×30×1) values of the attention identification data (one data of the V identification data DDf). The (30×30×1) values of the attention identification data are designated Vf _i , and the (30×30×1) random numbers Rf are designated Rf _i (i is an integer between 1 and (30×30×1) inclusive). In this case, each value Vf _i of the varying attention identification data is (Vf _i +Rf _i ).

In S340B, the CPU 210 calculates a fluctuation error value EVfb using the fluctuation attention identification data generated in S335B and the teacher data LDf. Specifically, the error value Egan calculated using the teacher data LDf values "0" as the teacher data values t _i in the above formula (1) and the fluctuation attention identification data values (Vf _i + Rf _i ) normalized by inputting them into a sigmoid function as a _i is the fluctuation error value EVfb.

In the first modified example, in S230 of FIG. 11, instead of the (2×V) fluctuation error values EVr and EVf, the (2×V) fluctuation error values EVrb and EVfb calculated in the fluctuation error value calculation process of FIG. 13 are used to adjust the multiple calculation parameters Pdn of the identification network DN.

In the first modified example described above, the varying step (S305, S310B, S330, S335B in FIG. 13) is a step of varying the values contained in the identification data DDr, DDf using the random numbers Rr, Rf. As a result, by varying the values contained in the identification data DDr, DDf, it is possible to adjust the progress of the training of the identification network DN with the same ease as in the embodiment.

Figure 14 is a flowchart of the variation error value calculation process of the second modified example. In the variation error value calculation process of Figure 14, steps S305C to S315C and S330C to S340C are executed instead of steps S305 to S315 and S330 to S340 of the variation error value calculation process of Figure 12.

In S305C, the CPU 210 calculates an error value Ev using the target identification data (one of the V pieces of identification data DDr) and the teacher data LDr. Specifically, the error value Egan calculated using the values "1" of the teacher data LDr as the values t _i of the teacher data in the above formula (1) and the values normalized by inputting the values of the target identification data into a sigmoid function as a _i is the error value Er.

In S310C, the CPU 210 obtains one random number Rrc. The random number Rrc is obtained independently from a predetermined normal distribution, for example.

In S315C, the CPU 210 adds the random number Rrc to the error value Er calculated in S305C to calculate the variation error value EVrc (EVrc = Er + Rrc).

In S330C, the CPU 210 calculates an error value Ef using the target identification data (one of the V pieces of identification data DDf) and the teacher data LDf. Specifically, the error value Ef is calculated by using the values "0" of the teacher data LDf as the values t _i of the teacher data in the above formula (1) and using the values normalized by inputting the values of the target identification data into a sigmoid function as a _i .

In S335C, the CPU 210 obtains one random number Rfc. The random number Rfc is obtained, for example, independently from the same normal distribution as the random number Rrc.

In S340C, the CPU 210 adds the random number Rfc to the error value Ef calculated in S330C to calculate the fluctuation error value EVfc (EVfc = Ef + Rfc).

In the second modified example, in S230 of FIG. 11, instead of the (2×V) fluctuation error values EVr and EVf, the (2×V) fluctuation error values EVrc and EVfc calculated in the fluctuation error value calculation process of FIG. 14 are used to adjust the multiple calculation parameters Pdn of the identification network DN.

In the second modified example described above, the variation process (S310C, S315C, S335C, S340C in FIG. 14) is a process of varying the error values Er and Ef calculated using the target identification data and the teacher data LDr and LDf, using random numbers Rrc and Rfc. The error values Er and Ef can be said to be index values calculated using a predetermined loss function (sigmoid cross entropy error in this embodiment). According to the second modified example, the variation error values EVrc and EVfc can be calculated using a smaller number of random numbers, so that the progress of training of the identification network DN can be easily adjusted.

(2) In the above embodiment and modified example, the values that are varied using random numbers are teacher data, identification data, and error values. Not limited to this, the CPU 210 may vary other values used to adjust the multiple calculation parameters Pdn of the identification network DN using random numbers. For example, the CPU 210 may vary intermediate parameters such as differential values calculated during the backpropagation method or gradient descent method using random numbers. Note that, if the values of these objects to be varied are allowed to take negative values, these values may be varied, for example, by adding random numbers that include negative values. Also, if the values of these objects to be varied are not allowed to take negative values, these values may be varied, for example, by adding random numbers that are 0 or greater.

In the above embodiment and modified example, a random number is added to the value of the object to be varied, thereby varying the value. Alternatively, for example, a random number may be obtained from a normal distribution with a mean of 1, and the value of the object to be varied may be varied by multiplying or dividing the value by the random number.

(3) In the above embodiment, in the network system 1000 constituting so-called GANs, the progress of training of the identification network DN is adjusted using random numbers so that the progress of training of the generation network GN does not lag too far behind the progress of training of the identification network DN. The present invention may be applied to various types of GANs, such as CycleGAN, DCGAN, and ConditionalGAN.

Furthermore, the present invention can be applied to cases other than GANs where, for some reason, it is necessary to adjust the progress of training of a specific machine learning model that is trained using output data and teacher data.

(4) In the above embodiment, as shown in formula (1), the sigmoid cross entropy error is used for the error values EVr, EVf, and Ef, and the mean absolute error is used for the error value Erf. Alternatively, other types of error values may be used for these error values. For example, the softmax cross entropy error may be used for the error values EVr, EVf, and Ef. The mean squared error may be used for the error value Erf.

(5) In the above embodiment, the generation network GN is generally a model that converts image data having a first attribute (e.g., an attribute of the character "A") and a second attribute (e.g., an attribute of the first font) and not having a third attribute (e.g., an attribute of the second font) into image data having the first attribute and the third attribute and not having the second attribute, as described above. Such conversion is generally also called "style transfer" or "image-to-image translation". As another specific example, the generation network GN may be a model that converts, for example, a line drawing (an image drawn only with lines) showing a predetermined object (still life such as a building or a shoe, or an animal such as a horse or a dog) into a color image showing the predetermined object. In addition, the generation network GN may be a model that converts, for example, an aerial photograph showing a predetermined place into a map showing the predetermined place. Even in these cases, the generative network GN can be said to be a model that converts image data that has a first attribute (an attribute indicating a specific object, an attribute indicating a specific place) and a second attribute (an attribute indicating a line drawing, an attribute indicating an aerial photograph) and does not have a third attribute (an attribute indicating a color image, an attribute indicating a map) into image data that has the first attribute and the third attribute and does not have the second attribute.

(6) In the above embodiment, in the generating network GN, both the input data ID and the fake data OD are image data. Not limited to this, both or either of the input data ID and the fake data OD may be data representing text or audio. For example, the generating network GN may be a model that converts image data representing a specific object into text or audio representing a specific object. Alternatively, the generating network GN may be a model that converts text representing a specific object into an image or audio representing a specific object.

(7) The specific configuration of the generating network GN and the identifying network DN is not limited to the configuration shown in FIG. 5 and FIG. 6, and may be various other configurations. For example, in the generating network GN and the identifying network DN, the number of layers of the convolution layer and the transposed convolution layer may be changed as appropriate. In addition, the generating network GN and the identifying network DN may include a pooling layer or a fully connected layer instead of all or a part of the convolution layer and the transposed convolution layer. In addition, various configurations are adopted for the post-processing performed on the values output from each layer. For example, the activation function used in the post-processing may be any other function, such as PReLU, softmax, or sigmoid. In addition, processes such as batch normalization and dropout may be added or omitted as post-processing as appropriate. Furthermore, in the generating network GN of the above embodiment, the m-th transposed convolution layer DL_m of the decoder DC is input with a value output from the (9-m)-th convolution layer EL_(9-m) of the encoder EC, but this input may be omitted.

The classification network DN may also be a machine learning model other than a neural network, such as a support vector machine (SVM), that is trained using teacher data.

(8) The hardware configuration of the data generating device 200 in FIG. 1 and the training device 100 in FIG. 7 is an example and is not limited to this. For example, the processor of the data generating device 200 and the training device 100 is not limited to a CPU, but may be a GPU (Graphics Processing Unit) or an ASIC (Application Specific Integrated Circuit), or a combination of these with a CPU. In addition, the training device 100 and the data generating device 200 may be multiple computers (for example, so-called cloud servers) that can communicate with each other via a network.

(9) In each of the above embodiments, a part of the configuration realized by hardware may be replaced by software, and conversely, a part or all of the configuration realized by software may be replaced by hardware. For example, the generation network GN and the identification network DN may be realized by a hardware circuit such as an ASIC (Application Specific Integrated Circuit) instead of a program module.

The present invention has been described above based on examples and modified examples, but the above-mentioned embodiments of the invention are intended to facilitate understanding of the present invention and do not limit the present invention. The present invention may be modified or improved without departing from the spirit and scope of the claims, and the present invention includes equivalents thereof.

100...Training device, 110...CPU, 120...Volatile storage device, 130...Non-volatile storage device, 140...Display unit, 150...Operation unit, 200...Data generation device, 210...CPU, 220...Volatile storage device, 230...Non-volatile storage device, 240...Display unit, 250...Operation unit, 270...Communication interface, 300...Printer, 1000...Network system, EL_0, L_0...Input layer, EL_1 to EL_8, L_0 to L_5...Convolution layer, DL_1 to DL_8...Transposed convolution layer, EC...Encoder, DC...Decoder, ID...Input data, OD...Output data (fake data), CD, CDf...Feature data, LDr, LDf...Teacher data, GN...Generative network, DN...Discrimination network, Pd, Pe, Pdn...Calculation parameters, Pr...Real data pair, PGg, PGt...Computer program

Claims (8)

1. A method for training a machine learning model, comprising:
A first step of inputting first input data into a first machine learning model to output first output data corresponding to the first input data, the first machine learning model being a model that performs a calculation process using a plurality of first calculation parameters on the first input data to output the first output data;
a second step of adjusting the plurality of first calculation parameters using the first output data and teacher data, the teacher data indicating a target value of the first output data corresponding to the first input data;
a third step of inputting second input data into a second machine learning model different from the first machine learning model, thereby outputting second output data corresponding to the second input data, the second machine learning model being a model that performs a calculation process using a plurality of second calculation parameters on the second input data to output the second output data;
a fourth step of adjusting the plurality of second calculation parameters using the second output data;
Equipped with
The first machine learning model and the second machine learning model are trained in parallel by a computer repeating the first step, the second step, the third step, and the fourth step a plurality of times;
the second step includes a varying step of varying values used for adjusting the plurality of first calculation parameters by using random numbers;
The method, wherein the fourth step does not include a step of varying values used to adjust the plurality of second operational parameters using random numbers. 2. The method of claim 1 ,
The varying step is a step of varying at least a part of values included in the plurality of teacher data by using the random numbers;
The method, wherein the second step is a step of adjusting the plurality of first calculation parameters using the first output data and the teacher data after at least a portion of the values have been varied. 2. The method of claim 1 ,
the varying step is a step of varying at least a part of values included in the first output data by using the random number;
The method, wherein the second step is a step of adjusting the plurality of first calculation parameters using the first output data after at least some of the values have been changed and the teacher data. 2. The method of claim 1 ,
The second step comprises:
Calculating an evaluation value based on a predetermined loss function using the first output data and the teacher data;
the varying step of varying the evaluation value by using the random number;
adjusting the plurality of first calculation parameters using the evaluation value after the change;
A method comprising: 2. The method of claim 1 ,
the first input data is a plurality of data including real data and the second output data as false data;
the teacher data is data indicating whether each of the plurality of first input data is the real data or the fake data,
the first machine learning model is a model for identifying whether each of the plurality of first input data is the real data or the fake data;
The first output data is data indicating a classification result by the first machine learning model,
the second machine learning model is a model for generating the second output data as the fake data,
In the second step, the plurality of first calculation parameters are adjusted using the first output data and the teacher data so that a difference between the first output data and the teacher data is reduced;
In the fourth step, the plurality of second calculation parameters are adjusted so that a difference between the first output data obtained by inputting the second output data as the fake data into the first machine learning model and the teacher data is increased. The method according to any one of claims 1 to 5,
The method includes varying values used for adjusting the plurality of first calculation parameters so that the values after the variation are equal to or greater than 0 in the varying step. A data generating device,
A generation unit that generates specific data using a learned second machine learning model trained by a training process using the first machine learning model;
an output unit that outputs the specific data;
Equipped with
the first machine learning model is a model that executes a calculation process using a plurality of first calculation parameters on first input data, and outputs identification data indicating a result of identifying whether the first input data is real data or fake data;
the second machine learning model is a model that performs a calculation process using a plurality of second calculation parameters on second input data to generate the fake data corresponding to the second input data;
The training process includes:
a fake data generation process of inputting the second input data into the second machine learning model to generate the fake data;
an identification data generation process for inputting a plurality of first input data, including real data and the fake data generated by the second machine learning model, into the first machine learning model and outputting a plurality of the identification data;
a first adjustment process for adjusting the plurality of first calculation parameters using the identification data and teacher data so that a difference between the identification data and the teacher data becomes small, the teacher data being data indicating whether each of the plurality of first input data is the real data or the fake data;
a second adjustment process for adjusting the plurality of second calculation parameters using the identification data and the teacher data obtained by inputting the false data into the first machine learning model so that a difference between the identification data and the teacher data becomes large;
a process of training the first machine learning model and the second machine learning model in parallel by repeating the fake data generation process, the identification data generation process, the first adjustment process, and the second adjustment process a plurality of times;
the first adjustment process includes a process of varying values used to adjust the plurality of first calculation parameters by using random numbers;
A data generating device, wherein the second adjustment process does not include a process of varying values used to adjust the plurality of second calculation parameters by using random numbers. A trained second machine learning model trained by a training process using a first machine learning model, the trained second machine learning model for generating specific data,
the first machine learning model is a model that executes a calculation process using a plurality of first calculation parameters on first input data, and outputs identification data indicating a result of identifying whether the first input data is real data or fake data;
the second machine learning model is a model that causes a computer to realize a function of executing a calculation process using a plurality of second calculation parameters on second input data to generate the fake data corresponding to the second input data;
The training process includes:
a fake data generation process of inputting the second input data into the second machine learning model to generate the fake data;
an identification data generation process for inputting a plurality of first input data, including real data and the fake data generated by the second machine learning model, into the first machine learning model and outputting a plurality of the identification data;
a first adjustment process for adjusting the plurality of first calculation parameters using the identification data and teacher data so that a difference between the identification data and the teacher data becomes small, the teacher data being data indicating whether each of the plurality of first input data is the real data or the fake data;
a second adjustment process for adjusting the plurality of second calculation parameters using the identification data and the teacher data obtained by inputting the false data into the first machine learning model so that a difference between the identification data and the teacher data becomes large;
a process of training the first machine learning model and the second machine learning model in parallel by repeating the fake data generation process, the identification data generation process, the first adjustment process, and the second adjustment process a plurality of times;
the first adjustment process includes a process of varying values used to adjust the plurality of first calculation parameters by using random numbers;
The second machine learning model that has been trained, wherein the second adjustment process does not include a process of varying values used to adjust the plurality of second calculation parameters using random numbers.

Images