Machine learning program, machine learning method, and information processing device

Abstract

An information processing device (100) acquires output data output from a trained autoencoder in response to the input of input data into the autoencoder. The information processing device (100) trains an autoencoder using a training data set including: output data; and data different from the data used for training the autoencoder.

Description

Machine learning program, machine learning method, and information processing device

[0001] The present invention relates to a machine learning program and the like.

[0002] Autoencoders are used for various tasks such as data dimensionality reduction, feature extraction, etc. Fig. 8 is a diagram for explaining an autoencoder.

[0003] 8, for example, an autoencoder 10 includes an encoder 10a and a decoder 10b. The encoder 10a encodes input data 11 to generate a low-dimensional feature representation 12. The decoder 10b generates reconstructed data 13 based on the feature representation 12 obtained by the encoder 10a.

[0004] The autoencoder 10 is trained to reduce the error between the input data 11 and the reconstructed data 13. This process gives the autoencoder 10 the ability to extract features from the data and generate similar data.

[0005] For example, if the input data 11 is an image (RGB image) with pn pixels, the input data 11 has N dimensions (N = pn × 3). Of the data represented in the N-dimensional space, only a small portion is meaningful as an image. If the input data 11 is input to the encoder 10a and converted into a feature representation 12, the image can be represented in an n-dimensional space, which is much smaller than the original N dimensions. The n-dimensional feature representation 12 is data that abstracts the input data 11, and this feature representation 12 can be used for other tasks such as classification.

[0006] By ensuring continuity between the space of input data and the space of the encoded result, the autoencoder 10 can be used as a generative model. For example, this generative model generates meaningful data in N dimensions by inputting values ​​into the encoded result. Conventional technologies such as VAE (Variational Autoencoder) have been proposed as such autoencoder-type generative models.

[0007] International Publication No. 2019 / 138655

[0008] When input data including an untrained pattern appears in the autoencoder 10, the autoencoder 10 is retrained. However, the conventional technology has a problem in that the cost of retraining is high.

[0009] For example, when retraining the autoencoder 10, the existing input data used in the previous training and input data including untrained patterns are used. The existing input data can be stored in the device itself or re-acquired from another device. Storing the existing input data in the device itself puts continuous strain on the storage device. Furthermore, even when re-acquiring the existing input data, it is assumed that the existing input data is stored in another device, which puts strain on the storage device of the other device.

[0010] Furthermore, if existing input data is not used and only input data containing untrained patterns is used, the system will no longer be able to process existing input data that it was previously able to handle (catastrophic forgetting).

[0011] In one aspect, the present invention aims to provide a machine learning program, a machine learning method, and an information processing device that can train an autoencoder without storing existing input data.

[0012] In a first proposal, a computer performs the following process: the computer obtains output data output from an autoencoder in response to input data being input to the trained autoencoder, and the computer trains the autoencoder using a training data set that includes the output data and data different from the data used to train the autoencoder.

[0013] It is possible to train an autoencoder without storing any existing input data.

[0014] FIG. 1 is a diagram illustrating the initial training of an autoencoder. FIG. 2 is a diagram illustrating an example of the data structure of statistical information. FIG. 3 is a diagram illustrating a process for generating data for retraining. FIG. 4 is a functional block diagram illustrating the configuration of an information processing device according to this embodiment. FIG. 5 is a flowchart illustrating the processing procedure of information processing according to this embodiment. FIG. 6 is a flowchart illustrating the processing procedure of generation processing. FIG. 7 is a diagram illustrating an example of the hardware configuration of a computer that realizes the same functions as the information processing device according to this embodiment. FIG. 8 is a diagram illustrating an autoencoder.

[0015] Hereinafter, embodiments of the machine learning program, the machine learning method, and the information processing device disclosed in the present application will be described in detail with reference to the accompanying drawings. However, the present invention is not limited to these embodiments.

[0016] The training performed by the information processing device according to the present embodiment on the autoencoder will be described below. Fig. 1 is a diagram for explaining the initial training of the autoencoder.

[0017] For the first training, the information processing device trains the autoencoder 10 using a training data table 50 prepared in advance. For example, the training data table 50 includes input image data 21-1, 21-2, 21-3, 21-4, ..., 21-m, where m is a natural number. Note that, although this embodiment describes the case where an image is used as input data, the present invention is also applicable to data other than images.

[0018] The information processing device obtains input image data 21-1 from the training data table 50 and inputs it to the encoder 10a to generate a feature representation 22-1. The information processing device inputs the feature representation 22-1 to the decoder 10b to generate reconstructed data 23-1. The information processing device trains the encoder 10a and the decoder 10b so that the error between the input image data 21-1 and the reconstructed data 23-1 is not small.

[0019] The information processing device also performs the same processing as above on the input image data 21-2, 21-3, 21-4, . . . , 21-m, thereby training the encoder 10a and the decoder 10b.

[0020] Here, in the process of performing the above training, the information processing device stores in the buffer 30 the feature representations 22-1, 22-2, 22-3, 22-4, ... 22-m that are generated when the input image data 21-1, 21-2, 21-3, 21-4, ... 21-m are input to the encoder 10a.

[0021] The information processing device calculates statistical information 35 based on the characteristic expressions 22-1 to 22-m stored in the buffer 30. For example, the information processing device calculates statistical quantities such as the average, variance, maximum value, and minimum value as the statistical information 35 based on the characteristic expressions 22-1 to 22-m.

[0022] 2 is a diagram showing an example of the data structure of the statistical information. As shown in FIG. 2, the statistical information sets the average, variance, maximum value, and minimum value for each element corresponding to each dimension. For example, if the feature expression has n dimensions, the information processing device calculates the statistics from the first element to the nth element and generates the statistical information 35.

[0023] The information processing device generates the number "m" of input image data 21-1 to 21-m stored in the training data table 50 as counter information 40. Note that the information processing device may generate the counter information 40 by counting up the value of the counter information every time input image data is input to the encoder 10a during the above training process.

[0024] When the initial training of the autoencoder 10 is completed, the information processing device empties the training data table 50 and deletes the feature representations 22-1 to 22-m stored in the buffer 30.

[0025] The information processing device performs various processes using the trained autoencoder 10, and when input image data including an untrained pattern appears, determines to perform retraining.

[0026] For example, the information processing device counts the number of times the difference between the input image data input to the encoder 10a and the reconstructed data output from the decoder 10b is equal to or greater than a threshold, and determines to perform retraining when the number of times is equal to or greater than a predetermined number. Furthermore, the information processing device registers the input image data when the difference between the input image data and the reconstructed data is equal to or greater than the threshold in the training data table 50 as input image data including an untrained pattern. In the following description, input image data including an untrained pattern will be referred to as "untrained input data."

[0027] When the information processing device determines to perform retraining, it generates data for the retraining. Fig. 3 is a diagram for explaining the process of generating data for the retraining. The information processing device generates a feature representation 36-1 based on the statistical information 35.

[0028] For example, the information processing device performs the following process to set the value of the first element of the feature representation 36-1. The information processing device acquires statistics for the first element set in the statistics information 35, sets a normal distribution based on the mean and variance included in the statistics, and generates the value of the first element using random number generation according to the normal distribution. The Box-Muller algorithm may be used to generate random numbers according to the normal distribution. The information processing device may also adjust the value of the first element so that it is included in the range from the minimum value to the maximum value of the statistics. The information processing device performs similar processes for the other elements to generate the feature representation 36-1.

[0029] The information processing device generates reconstructed data 37-1 by inputting the feature representation 36-1 to the decoder 10b. The information processing device registers the reconstructed data in the training data table 50 as input image data for training.

[0030] The information processing device generates feature representations based on the statistical information 35, and repeats a series of processes for generating reconstructed data the number of times set in the counter information 40. For example, if "m" is registered in the counter information 40, the information processing device repeats the above process m times to generate m pieces of reconstructed data and register them in the training data table 50. This makes it possible to generate similar input image data by using the statistical information 35, even if the input image data used in the previous training is not saved in the training data table 50. Note that the training data table 50 includes not only the reconstructed data generated by the above process, but also untrained input data.

[0031] The information processing device performs retraining of the autoencoder 10 using the training data table 50 described in Fig. 3. The retraining performed by the information processing device is similar to the initial training described in Fig. 1. In the process of performing retraining, the information processing device stores in the buffer 30 the feature representations generated when input image data (reconstructed data or untrained input data) in the training data table 50 is input to the encoder 10a, calculates statistics for each element, and updates the statistics information 35.

[0032] The information processing device also updates the counter information 40. For example, if m pieces of reconstructed data and l pieces of untrained input data are registered in the training data table 50, the value set in the counter information 40 will be "m+l".

[0033] When the retraining is completed, the information processing device empties the training data table 50 and deletes the feature representations stored in the buffer 30. The information processing device performs various processes using the trained autoencoder 10, and when input image data including an untrained pattern appears, the information processing device determines to perform retraining. When the information processing device determines to perform retraining, it performs the above processes again.

[0034] As described above, when the information processing device according to this embodiment initially trains the autoencoder 10, it generates the statistical information 35 and the counter information 40 based on each feature representation obtained by inputting each input image data in the training data table 50 into the encoder 10a. When training is completed, the information processing device empties the training data table 50 and the buffer 30.

[0035] The information processing device inputs new input image data to a trained autoencoder 10, and when the number of times that the difference between the reconstructed data output from the autoencoder 10 and the input image data is equal to or greater than a threshold is equal to or greater than a predetermined number of times, the information processing device determines to perform retraining of the autoencoder 10. When the difference between the reconstructed data and the input image data is equal to or greater than the threshold, the information processing device performs a process of registering the input image data in the training data table 50.

[0036] When retraining the autoencoder 10, the information processing device generates a feature representation based on the statistical information 35 and inputs the feature representation to the decoder 10b to generate reconstructed data. The information processing device executes the process of generating reconstructed data the number of times set in the counter information 40, thereby registering data corresponding to the input image data used in the previous training and the same number of data as the input image data used in the previous training in the training data table 50. The information processing device retrains the autoencoder 10 using the training data table 50. This makes it possible to train the autoencoder without storing existing input data.

[0037] Next, an example of the configuration of an information processing device that executes the processes described in Figures 1 and 3 will be described. Figure 4 is a functional block diagram showing the configuration of an information processing device according to this embodiment. As shown in Figure 4, this information processing device 100 has a communication unit 110, an input unit 120, a display unit 130, a storage unit 140, and a control unit 150.

[0038] The communication unit 110 performs data communication with an external device or the like via a network. The communication unit 110 is a network interface card (NIC) or the like. For example, the communication unit 110 acquires input image data or the like from the external device when training for the first time.

[0039] The input unit 120 is an input device that inputs various types of information to the control unit 150 of the information processing device 100. For example, the input unit 120 corresponds to a keyboard, a mouse, a touch panel, etc. The user may operate the input unit 120 to give an instruction to perform the first training.

[0040] The display unit 130 is a display device that displays information output from the control unit 150 .

[0041] The storage unit 140 includes a buffer 30, statistical information 35, counter information 40, a training data table 50, and model information 60. The storage unit 140 is a storage device such as a memory.

[0042] The buffer 30 temporarily stores the feature representations output from the encoder 10a when training the autoencoder 10. The feature representations correspond to "output values."

[0043] The statistical information 35 is information generated during training. As described in Fig. 2, the statistical information 35 includes statistics of the mean, variance, maximum value, and minimum value set for each element corresponding to each dimension.

[0044] The counter information 40 is information generated during training. The counter information 40 contains the number of input image data input to the encoder 10a of the autoencoder 10 during training.

[0045] The training data table 50 stores data for training the autoencoder 10. For the first training, the input image data 21-1 to 21-m prepared in advance are stored. After the first training is completed, the input image data 21-1 to 21-m are deleted from the training data table 50.

[0046] After the first training is completed, the training data table 50 stores untrained input data. The training data table 50 also stores reconstructed data generated by the process described in FIG. 3. Every time training (first training) is completed, the data in the training data table 50 is deleted. Each piece of data stored in the training data table 50 corresponds to a "training data set."

[0047] The model information 60 holds information about the autoencoder 10. The autoencoder 10 has an encoder 10a and a decoder 10b. The encoder 10a generates a feature representation by encoding input data. The decoder 10b generates reconstructed data based on the feature representation obtained by the encoder 10a. The input data and the reconstructed data have N dimensions. The feature representation has n dimensions (N>n). The autoencoder 10 corresponds to an "autoencoder."

[0048] The control unit 150 includes an acquisition unit 151, a learning processing unit 152, a determination unit 153, and a generation unit 154. The control unit 150 is a central processing unit (CPU), a graphics processing unit (GPU), or the like.

[0049] The acquisition unit 151 acquires various types of information from an external device or the like via the communication unit 110. For example, the acquisition unit 151 acquires input image data for the first training from the external device and stores the data in the training data table 50.

[0050] The learning processing unit 152 uses the training data table 50 to train (retrain) the autoencoder 10. As described in FIG. 1, during the initial training, the training data table 50 stores input image data 21-1 to 21-m. As described in FIG. 3, during retraining, the training data table 50 stores reconstructed data and untrained input data. In the description of the learning processing unit 152, for convenience of explanation, the input image data, reconstructed data, and untrained input data stored in the training data table 50 will be collectively referred to as input data.

[0051] The learning processing unit 152 acquires input data from the training data table 50 and inputs it to the encoder 10a to generate feature representations. The learning processing unit 152 inputs the feature representations to the decoder 10b to generate reconstructed data. The learning processing unit 152 trains the encoder 10a and the decoder 10b so that the error between the input data and the reconstructed data is not small.

[0052] During the training process, the learning processing unit 152 stores the feature representations generated when input data is input to the encoder 10a in the buffer 30. The learning processing unit 152 generates statistical information 35 based on the feature representations stored in the buffer 30 and registers it in the storage unit 140. If the statistical information 35 generated during the previous training is registered in the storage unit 140, the learning processing unit 152 updates it with the statistical information 35 generated this time.

[0053] The learning processing unit 152 generates counter information 40 during the training process and registers the counter information 40 in the storage unit 140. If the counter information 40 generated during the previous training is registered in the storage unit 140, the learning processing unit 152 updates it with the counter information 40 generated this time.

[0054] When the training (retraining) of the autoencoder 10 is completed, the learning processing unit 152 empties the training data table 50 and deletes the feature representations stored in the buffer 30. Other processing related to the learning processing unit 152 is similar to the processing described in Figures 1, 3, etc.

[0055] The learning processing unit 152 executes initial training on the autoencoder 10 when receiving an instruction to execute initial training from the input unit 120. The learning processing unit 152 executes retraining on the autoencoder 10 when receiving a request to execute retraining from the determination unit 153, which will be described later.

[0056] The determination unit 153 acquires new input image data to be processed from an external device or the like via the communication unit 110, and determines whether or not to retrain the autoencoder 10 using the input image data.

[0057] The determination unit 153 inputs input image data to the encoder 10a and acquires reconstructed data from the decoder 10b. The determination unit 153 counts the number of times the difference between the input image data input to the encoder 10a and the reconstructed data output from the decoder 10b is equal to or greater than a threshold, and when this number is equal to or greater than a predetermined number, determines to execute retraining and outputs a request to execute retraining to the learning processing unit 152. Furthermore, when the determination unit 153 determines to execute retraining, it outputs a request to generate data to the generation unit 154.

[0058] The determination unit 153 registers input image data (untrained input data) in the training data table 50 when the difference between the input image data and the reconstructed data is equal to or greater than a threshold value.

[0059] When the generation unit 154 receives a data generation request, it generates a plurality of feature representations based on the statistical information 35. The generation unit 154 generates a plurality of reconstructed data from the plurality of feature representations, and registers the plurality of reconstructed data in the training data table 50 as data to be used during training.

[0060] For example, the generation unit 154 performs the following process to set the value of the first element of the feature expression. The generation unit 154 acquires the statistics of the first element set in the statistics information 35, sets a normal distribution based on the mean and variance included in the statistics, and generates the value of the first element using random number generation according to the normal distribution. The Box-Muller algorithm may be used to generate random numbers according to the normal distribution. The generation unit 154 may also adjust the value of the first element so that the value is included in the range from the minimum value to the maximum value of the statistics. The generation unit 154 performs similar processes for the other elements to generate feature expressions.

[0061] The generation unit 154 generates feature representations based on the statistical information 35, and repeats a series of processes for generating reconstructed data the number of times set in the counter information 40, and stores each reconstructed data in the training data table 50. The processing by the generation unit 154 corresponds to the processing described in FIG. 3 .

[0062] Next, an example of the processing procedure of the information processing device 100 according to this embodiment will be described. Fig. 5 is a flowchart showing the processing procedure of the information processing procedure according to this embodiment. As shown in Fig. 5, the learning processing unit 152 of the information processing device 100 performs initial training on the autoencoder 10 based on input image data stored in the training data table 50 (step S101).

[0063] The learning processing unit 152 generates the statistical information 35 and the counter information 40 (step S102). The learning processing unit 152 deletes the information stored in the training data table 50 and the information stored in the buffer 30 (step S103).

[0064] The acquisition unit 151 of the information processing device 100 acquires new input image data from an external device or the like (step S104). The determination unit 153 of the information processing device 100 inputs the input image data to the autoencoder 10 and generates reconstructed data (step S105).

[0065] If the difference between the input image data and the reconstructed data is equal to or greater than a threshold, the determination unit 153 registers the input image data (untrained input data) in the training data table 50 (step S106). The determination unit 153 determines whether the number of times the difference between the input image data and the reconstructed data is equal to or greater than the threshold is equal to or greater than a threshold (step S107).

[0066] If the number of times is not equal to or greater than the predetermined number (No at step S108), the determination unit 153 proceeds to the process at step S112.

[0067] On the other hand, if the number of times is equal to or greater than the predetermined number (Yes at step S108), the determination unit 153 proceeds to the process at step S109. The generation unit 154 of the information processing device 100 executes the generation process (step S109).

[0068] The learning processing unit 152 performs training on the autoencoder 10 based on the input data (untrained input data, reconstructed data) stored in the training data table 50 (step S110). The learning processing unit 152 updates the statistics information 35 and the counter information 40 (step S111).

[0069] If the information processing device 100 determines to continue the process (Yes at step S112), the process proceeds to step S103. On the other hand, if the information processing device 100 determines not to continue the process (No at step S112), the process ends.

[0070] Next, an example of the generation process described in step S109 of Fig. 5 will be described. Fig. 6 is a flowchart showing the processing procedure of the generation process. As shown in Fig. 6, the generation unit 154 of the information processing device 100 sets i = 0 (step S201). The generation unit 154 generates a feature representation based on the statistical information 35 (step S202).

[0071] The generation unit 154 inputs the feature representation to the decoder 10b to generate reconstructed data (step S203). The generation unit 154 registers the reconstructed data in the training data table 50 (step S204). The generation unit 154 sets i=i+1 (step S205).

[0072] If the value of i is not equal to the value of the counter information 40 (step S206, No), the generation unit 154 proceeds to step S202. If the value of i is equal to the value of the counter information 40 (step S206, Yes), the generation unit 154 ends the process.

[0073] Next, the effects of the information processing device 100 according to this embodiment will be described. When retraining the autoencoder 10, the information processing device 100 generates feature representations based on the statistical information 35, inputs the feature representations to the decoder 10b, thereby generating reconstructed data, and registers the reconstructed data in the training data table 50. The reconstructed data registered in the training data table 50 can be said to correspond to the input image data used in the previous training. In other words, the information processing device 100 can train the autoencoder without storing existing input data.

[0074] The information processing device 100 generates a plurality of feature representations by inputting each piece of input image data registered in the training data table 50 to the encoder 10a, and generates statistical information 35 based on the plurality of feature representations. This makes it possible to extract the statistics of the input image data stored in the training data table 50.

[0075] The information processing device 100 executes the process of generating reconstruction data using statistical information 35 the number of times set in counter information 40, thereby being able to register in the training data table 50 data that corresponds to the input image data used during the previous training and is the same number of data as the input image data used during the previous training.

[0076] The information processing device 100 determines whether to retrain the autoencoder 10 based on the difference between the input image data input to the autoencoder 10 and the reconstructed data output from the autoencoder 10. This makes it possible to detect the appearance of input image data that includes an untrained pattern.

[0077] Next, a description will be given of an example of the hardware configuration of a computer that realizes the same functions as the above-described information processing device 100. Fig. 7 is a diagram showing an example of the hardware configuration of a computer that realizes the same functions as the information processing device of this embodiment.

[0078] 7, computer 300 has a CPU 301 that executes various types of arithmetic processing, an input device 302 that accepts data input from a user, and a display 303. Computer 300 also has a communication device 304 that exchanges data with external devices via a wired or wireless network, and an interface device 305. Computer 300 also has a RAM 306 that temporarily stores various types of information, and a hard disk drive 307. Each of devices 301 to 307 is connected to a bus 308.

[0079] The hard disk drive 307 stores an acquisition program 307 a, a learning processing program 307 b, a determination program 307 c, and a generation program 307 d. The CPU 301 reads out each of the programs 307 a to 307 d and loads them into the RAM 306.

[0080] The acquisition program 307a functions as an acquisition process 306a. The learning processing program 307b functions as a learning processing process 306b. The determination program 307c functions as a determination process 306c. The generation program 307d functions as a generation process 306d.

[0081] The processing of the acquisition process 306a corresponds to the processing of the acquisition unit 151. The processing of the learning process 306b corresponds to the processing of the learning processing unit 152. The processing of the determination process 306c corresponds to the processing of the determination unit 153. The processing of the generation process 306d corresponds to the processing of the generation unit 154.

[0082] It should be noted that each of the programs 307a to 307d does not necessarily have to be stored in the hard disk drive 307 from the beginning. For example, each program may be stored in a "portable physical medium" such as a flexible disk (FD), CD-ROM, DVD, magneto-optical disk, or IC card that is inserted into the computer 300. Then, the computer 300 may read and execute each of the programs 307a to 307d.

[0083] 30 Buffer 35 Statistical quantity information 40 Counter information 50 Training data table 60 Model information 100 Information processing device 110 Communication unit 120 Input unit 130 Display unit 140 Storage unit 150 Control unit 151 Acquisition unit 152 Learning processing unit 153 Determination unit 154 Generation unit

Claims

1. A machine learning program that causes a computer to execute a process of obtaining output data output from a trained autoencoder in response to input data being input to the autoencoder, and training the autoencoder using a training data set that includes the output data and data different from the data used to train the autoencoder.

2. The machine learning program of claim 1, wherein the autoencoder has an encoder and a decoder, and further causes a computer to execute a process of obtaining multiple output values ​​output from the encoder by inputting multiple data included in the training dataset into the encoder, and generating statistical information based on the multiple output values, and the process of obtaining the output data comprises generating a value based on the statistical information and inputting the generated value into the decoder to obtain the output data.

3. The machine learning program of claim 2, further comprising a process of counting the number of data input to the autoencoder when training the autoencoder using a plurality of data included in the training dataset, and wherein the process of obtaining the output data comprises generating a number of values ​​equal to the counted number and inputting each of the generated values ​​to the decoder to obtain the output data.

4. The machine learning program of claim 1, further comprising causing the computer to execute a process of determining whether or not to retrain the autoencoder based on the difference between input data input to the autoencoder and output data output from the autoencoder.

5. A machine learning method, characterized in that a computer executes a process of obtaining output data output from a trained autoencoder in response to input data being input to the autoencoder, and training the autoencoder using a training data set including the output data and data different from the data used to train the autoencoder.

6. The machine learning method of claim 5, wherein the autoencoder has an encoder and a decoder, and the computer further executes a process of obtaining a plurality of output values ​​output from the encoder by inputting a plurality of data included in the training dataset into the encoder, and generating statistical information based on the plurality of output values, and the process of obtaining the output data includes generating a value based on the statistical information and inputting the generated value into the decoder to obtain the output data.

7. The machine learning method of claim 6, further comprising the step of causing a computer to execute a process of counting the number of data input to the autoencoder when training the autoencoder using a plurality of data included in the training dataset, and the process of obtaining the output data comprises generating a number of values ​​equal to the counted number and inputting each of the generated values ​​to the decoder to obtain the output data.

8. The machine learning method according to claim 5, further comprising causing the computer to execute a process of determining whether or not to retrain the autoencoder based on the difference between input data input to the autoencoder and output data output from the autoencoder.

9. An information processing device having a control unit that executes a process of acquiring output data output from a trained autoencoder in response to input data being input to the autoencoder, and training the autoencoder using a training data set including the output data and data different from the data used to train the autoencoder.

10. The information processing device according to claim 9, characterized in that the control unit further executes a process in which the autoencoder has an encoder and a decoder, and obtains a plurality of output values ​​output from the encoder by inputting a plurality of data included in the training data set into the encoder, and generates statistical information based on the plurality of output values, and the process of obtaining the output data generates a value based on the statistical information and inputs the generated value into the decoder to obtain the output data.

11. The information processing device described in claim 10, characterized in that the control unit further executes a process of counting the number of data input to the autoencoder when training the autoencoder using multiple data included in the training dataset, and the process of obtaining the output data generates a number of values ​​equal to the counted number and inputs each of the generated values ​​to the decoder to obtain the output data.

12. The information processing device according to claim 9, characterized in that the control unit further executes a process of determining whether or not to retrain the autoencoder based on the difference between input data input to the autoencoder and output data output from the autoencoder.

Images