JP2023000929A - Machine learning device and machine learning method - Google Patents

Abstract

To create a highly accurate learning model even when using composite data.SOLUTION: A machine learning device stores a first learning model for outputting a feature quantity of input data, a second learning model for inputting the feature quantity output from the first learning model to output an identification error whether or not the data input to the first learning model is actual data, and a third learning model for inputting the identification error output from the second learning model to output generated data obtained by imparting features of the actual data to composite data. The machine learning device inputs the actual data, the composite data, and the generated data to the first learning model and performs machine learning of the second learning model on the basis of the output feature quantity. The machine learning device inputs the feature quantity output from the first learning model to the second learning model and performs machine learning of the third learning model on the basis of the output identification error. The machine learning device input the identification error output from the second learning model to the third learning model and performs machine learning of the first learning model on the basis of the output generated data.SELECTED DRAWING: Figure 7

Description

TECHNICAL FIELD Embodiments of the present invention relate to a machine learning device and a machine learning method.

2. Description of the Related Art In recent years, much research and development has been carried out on learning models that input data and output feature amounts (intermediate output data) and final output data of the data. According to such a learning model, for example, it is possible to detect people, vehicles, and the like from image data. In addition to real data, synthetic data may also be used as data for learning a learning model. Then, for example, even if there is little actual data, it is possible to increase the amount of learning data by using synthetic data.

JP 2018-163444 A JP-A-2021-12595 WO2018/198233

However, in the prior art, synthetic data may have characteristics different from those of actual data, and learning models created using synthetic data are often not highly accurate.

Therefore, the present invention has been made in view of the above circumstances, and provides a machine learning device and a machine learning method that can stably create a highly accurate learning model even when synthetic data is used. The task is to

The machine learning device of the embodiment includes, for example, a first learning model that inputs data and outputs a feature amount of the data, and a first learning model that inputs the feature amount output from the first learning model. a second learning model that outputs a discrimination error as to whether or not the data input to is real data; a storage unit that stores a third learning model that outputs generated data to which features are added; a second learning processing unit that performs a second process of machine learning the second learning model based on the feature amount; and inputting the feature amount output from the first learning model to the second learning model. a third learning processing unit that performs a third process of machine-learning the third learning model based on the output identification error; a first learning processing unit that inputs an error and performs a first process of machine learning the first learning model based on the output generated data.
According to this configuration, for example, the third learning model outputs the generated data in which the features of the real data are added to the synthetic data, and by using the generated data, the highly accurate first learning model is stabilized. can be created by

Further, the machine learning device of the embodiment repeats the second process and the third process until a first repetition end condition is satisfied, and then performs the first process. The process is repeated until the second iteration end condition is met.
According to this configuration, the accuracy of the first learning model can be further improved by, for example, repeating the entire above-described processing.

Further, the machine learning device of the embodiment further includes, for example, a difference calculation processing unit that calculates a difference between the synthesized data and the generated data, and the third learning processing unit further uses the difference to calculate the third Perform the process of 3.
According to this configuration, for example, by using the above-described difference, it is possible to avoid a situation where the third learning processing unit generates generated data that is far from the synthesized data just to bring the synthesized data closer to the real data.

Further, in the machine learning device of the embodiment, for example, the third learning processing unit weights the identification error and the difference and performs the third processing.
According to this configuration, it is possible to further improve the accuracy of the first learning model and shorten the computation processing time by weighting according to the scene, such as changing the weights in the first half and the second half of the entire repeated processing.

Further, the machine learning method of the embodiment includes, for example, a first learning model that inputs data and outputs a feature amount of the data, and a first learning model that inputs the feature amount output from the first learning model. a second learning model that outputs a discrimination error indicating whether or not data input to the learning model is real data; and a third learning model that outputs generated data to which data features are added, wherein the actual data, the synthetic data, and the generated data are input to the first learning model. a second learning processing step of performing a second process of machine-learning the second learning model based on the output feature amount; and providing the second learning model with the feature output from the first learning model a third learning processing step of performing a third process of inputting a quantity and performing machine learning of the third learning model based on the output identification error; a first learning processing step of inputting the outputted identification error and performing a first processing of machine learning the first learning model based on the outputted generated data.
According to this configuration, for example, the third learning model outputs the generated data in which the features of the real data are added to the synthetic data, and by using the generated data, the highly accurate first learning model is stabilized. can be created by

FIG. 1 is a diagram illustrating an example of a hardware configuration of a machine learning device according to an embodiment; FIG. 2 is a diagram illustrating a functional configuration of the machine learning device according to the embodiment; FIG. 3 is a diagram illustrating examples of synthesized data and generated data according to the embodiment. FIG. 4 is an explanatory diagram of learning of the second learning model in the embodiment. FIG. 5 is an explanatory diagram of learning of the third learning model in the embodiment. FIG. 6 is an explanatory diagram of learning of the first learning model in the embodiment. FIG. 7 is a flowchart illustrating processing executed by the machine learning device according to the embodiment;

Illustrative embodiments of the invention are disclosed below. The configurations of the embodiments shown below and the actions, results, and effects brought about by the configurations are examples. The present invention can be realized by configurations other than those disclosed in the following embodiments, and at least one of various effects based on the basic configuration and derivative effects can be obtained. .

In order to facilitate understanding of the present embodiment, the related art will be explained again. Machine learning and learning models are hereinafter also referred to as AI (Artificial Intelligence).

In general, there is a problem that when AI is trained on synthetic data, it does not match real data. It is believed that this is because there is a large gap between synthetic data and real data in feature amounts calculated by AI. As methods for filling this gap, for example, the following methods 1 to 3 have been proposed so far.

In method 1 (Patent Document 1), a discriminator that classifies CG (Computer Graphics) data and real-world data is used to correct CG data so that the CG data and real-world data are classified, so that CG data and real-world data are classified. Fill gaps between data.

Also, in Method 2 (Patent Document 2), the distance of the distribution of the feature amount of the live-action data/CG data calculated by the AI to be learned is measured, and by making the AI learn with a constraint that the distance becomes small, To fill the feature amount gap between actual data/CG data.

In addition, in method 3 (Patent Document 3), a mechanism for identifying the gap (offset) between the feature amount of CG data and actual data is incorporated into AI, and by learning AI with the feature amount after subtracting the offset, CG data / Filling the gaps in live-action data.

Methods 1 to 3 are techniques for filling the gap between CG data/actual data, but there are problems. For example, in the method of making CG data look like real-life data from a human perspective, it is based on the idea that since it looks like real-life data to humans, it will also be close to real-life data for AI. However, AI does not always recognize images with the same characteristics as humans. Therefore, AI that actually learns from CG data cannot know whether the CG data is close to the actual data. In other words, it is not always possible for AI to learn effectively with CG data.

In addition, in the method of incorporating a mechanism that fills the gap between CG data / live-action data into AI that learns with CG data, AI learns so that it is suitable for both live-action data / CG data (so that it can calculate the feature amount). , If you want to improve the performance of AI that has learned only with live-action data (improving the performance of existing AI), using this method may reduce the suitability for live-action data by trying to adapt to CG data. Occur.

In other words, the synthetic data may have different characteristics from the actual data, and in the conventional technology, the accuracy of the learning model created using the synthetic data is often not high.

Therefore, a technique for stably creating a highly accurate learning model even when synthetic data is used will be described below.

FIG. 1 is a diagram showing an example of a hardware configuration of a machine learning device 100. As shown in FIG. As shown in FIG. 1, machine learning device 100 includes processor 101, ROM 102, RAM 103, input unit 104, display unit 105, communication I/F 106, and HDD 107. In this example, machine learning device 100 has a hardware configuration similar to that of a normal computer. Note that the hardware elements of the machine learning device 100 are not limited to the hardware elements illustrated in FIG.

The processor 101 is a hardware circuit including, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an MPU (Micro processing unit), and the like. By executing a program, the processor 101 comprehensively controls the operation of the machine learning device 100 and implements various functions of the machine learning device 100 . Various functions of the machine learning device 100 will be described later.

The ROM 102 is a non-volatile memory and stores various data including programs for activating the machine learning device 100 . A RAM 103 is a volatile memory having a work area for the processor 101 .

The input unit 104 is a device for a user using the machine learning device 100 to perform various operations. The input unit 104 is composed of, for example, a mouse, keyboard, touch panel, hardware keys, and the like.

The display unit 105 displays various information. The display unit 105 is configured by, for example, a liquid crystal display, an organic EL (Electro Luminescence) display, or the like. Note that, for example, the input unit 104 and the display unit 105 may be configured integrally in a form like a touch panel. Communication I/F 106 is an interface for connecting with a network. A HDD (Hard Disk Drive) 107 stores various data.

FIG. 2 is a diagram showing the functional configuration of the machine learning device 100 of the embodiment. A machine learning device 100 includes a processing unit 1 and a storage unit 2 . The storage unit 2 is realized by the ROM 102, the RAM 103, and the HDD 107, for example. The storage unit 2 stores various data. The storage unit 2 stores a first learning model 21, a second learning model 22, and a third learning model 23, for example. Each learning model is data used by the processing unit 1, but for convenience of explanation, the learning model may be described as inputting or outputting.

Further, among the first learning model 21, the second learning model 22, and the third learning model 23, the first learning model 21 is the direct target for improving the accuracy. The second learning model 22 and the third learning model 23 are used to improve the accuracy of the first learning model 21 . Further, hereinafter, actual data is actually obtained data, for example, photographed data. Synthetic data is data synthesized as a substitute for real data, such as CG data. The generated data is data obtained by adding the characteristics of the actual data to the synthesized data (data obtained by processing the synthesized data).

Here, FIG. 3 is a diagram showing an example of (a) synthesized data and (b) generated data in the embodiment. The generated data shown in FIG. 3B is obtained by adding the features of the real data to the synthetic data shown in FIG. 3A. In FIG. 3, the only difference between the two data is the presence or absence of the shadow of the vehicle C on the road R, but the present invention is not limited to this. For example, the vehicle C portion may be processed more complicatedly, and the road R other than the vehicle C may be processed.

Returning to FIG. 2, the first learning model 21 inputs data (real data, synthetic data, real data) and outputs data feature amounts (intermediate output data) and final output data. The type of machine learning including the first learning model 21 is arbitrary.

The second learning model 22 inputs the feature amount (intermediate output data) and final output data output from the first learning model 21, and determines whether the data input to the first learning model 21 is real data. identify and output the identification error. In addition, the second learning model 22 acquires correct data indicating whether the data input to the first learning model 21 is real data (for example, a label indicating a distinction between real data, synthetic data, and generated data). Thus, it is possible to determine whether or not the identification result is correct, and to calculate the identification error.

The third learning model 23 receives the identification error output from the second learning model 22 and outputs generated data obtained by adding the features of the real data to the synthetic data.

The processing unit 1 is implemented by the processor 101 executing programs stored in the ROM 102 and the HDD 107 . The processing unit 1 executes various arithmetic processing. The processing unit 1 includes, as a functional configuration, a first learning processing unit 11, a second learning processing unit 12, a third learning processing unit 13, an estimation processing unit 14, a setting unit 15, a difference calculation processing unit 16, and a control unit 17. .

Here, FIG. 4 is an explanatory diagram of learning of the second learning model 22 in the embodiment. The second learning processing unit 12 inputs real data, synthetic data, and generated data to the first learning model 21, and based on the output feature amount (intermediate output data) and final output data, the second learning model 22 is machine-learned. Since there is no generated data at the first time, the second processing is performed using the actual data and the synthetic data.

FIG. 5 is an explanatory diagram of learning of the third learning model 23 in the embodiment. The third learning processing unit 13 inputs the feature amount (intermediate output data) and the final output data output from the first learning model 21 to the second learning model 22, and performs third learning based on the output identification error. A third process of machine-learning the learning model 23 is performed.

Also, the difference calculation processing unit 16 calculates the difference between the synthesized data and the generated data. Then, the third learning processing unit 13 further uses the difference to perform the third processing. Also, the third learning processing unit 13 may perform the third processing by weighting the identification error and the difference. The third learning processing section 13 will be described in more detail below.

For example, when the identification error (0 to 1) by the second learning model 22 is e, the third learning processing unit 13 sets (1-e) as the error of the third learning model 23, and the third learning model 23 feedback to In other words, the third learning processing unit 13 learns how to convert the synthesized data so that the error of the third learning model 23 is reduced. At this time, based on the difference between the synthesized data and the generated data calculated by the difference calculation processor 16, the third learning processing unit 13 prevents the synthesized data and the generated data from being separated from each other. As a result, the third learning processing unit 13 generates generated data such that the first learning model 21 is deceived as real data. Further, the third learning processing unit 13 may weight the identification error and the difference when learning the third learning model 23 . For example, the weight of the identification error of the second learning model 22 may be increased at the beginning of learning, and the weight of the difference between the synthesized data/generated data may be increased when the first learning model 21 is deceived to some extent.

FIG. 6 is an explanatory diagram of learning of the first learning model 21 in the embodiment. The first learning processing unit 11 inputs the identification error output from the second learning model 22 to the third learning model 23, and machine-learns the first learning model 21 based on the output generated data. process. In this machine learning, actual data is also input to the first learning model 21 .

Returning to FIG. 2, the processing unit 1 repeats, for example, the second processing and the third processing until the first repetition end condition is satisfied, and then performs the first processing. The process is repeated until the second repetition end condition is satisfied (details will be described later with reference to FIG. 7).

The estimation processing unit 14 performs estimation processing such as object detection by inputting data to the first learning model 21 that has completed learning and outputting the feature amount (intermediate output data) and final output data of the data. .

The setting unit 15 sets various parameters, weighting, and the like.

The control unit 17 performs processing other than the processing by each unit 11-16. The control unit 17 controls display of various information on the display unit 105, for example.

Next, processing executed by the machine learning device 100 will be described. FIG. 7 is a flowchart showing processing executed by the machine learning device 100 of the embodiment.

First, in step S1, the second learning processing unit 12 inputs real data and synthetic data to the first learning model 21, and performs a second learning process based on the output feature amount (intermediate output data) and final output data. A second process of machine-learning the learning model 22 is performed (FIG. 4).

Next, in step S2, the second learning processing unit 12 determines whether or not learning is completed, that is, whether or not a predetermined learning end condition is satisfied. Return to step S1. As the learning termination condition here, for example, a condition provided for a loss during learning of the second learning model 22 can be considered. More specifically, the end of learning is generally judged by the balance between the loss for training data and the loss for validation data, but it is also possible to use a mechanism that automatically judges the completion of learning, such as early stopping. good.

In step S3, the third learning processing unit 13 inputs the feature amount (intermediate output data) and the final output data output from the first learning model 21 to the second learning model 22, and converts the output identification error into Based on this, a third process of machine learning the third learning model 23 is performed.

Next, in step S4, the third learning processing unit 13 determines whether or not learning is completed, that is, whether or not a predetermined learning end condition is satisfied. Return to step S3. Note that, as the learning termination condition here, for example, the identification error of the second learning model 22, the number of times of learning, etc. can be considered. More specifically, the learning ends here, for example, when the second learning model 22 can no longer correctly distinguish between real data and generated data (when the rate of correct discrimination becomes around 0.5). be. Considering the case where the number of times of learning is about 0.5 by chance at the beginning of learning, the learning end condition may be that the number of times of learning is a certain number.

In step S5, the processing section 1 determines whether or not the first repetition end condition is satisfied. If Yes, the process proceeds to step S6, and if No, the process returns to step S1. As the first repetition end condition, for example, the number of repetitions (N1) can be considered.

Using generated data that has been refined by repeating steps S1 to S4 (actualization processing has progressed), the first learning model 21 that aims to improve performance by utilizing synthetic data is learned from step S6 onwards. For example, if the object detection AI (the first learning model 21) is to be learned, information on the type and position of the object to be detected is given as teacher data, and learning is performed so that the object can be detected correctly. This teacher data is created when synthetic data is created. Steps after step S6 will be described below.

In step S6, the first learning processing unit 11 inputs the identification error output from the second learning model 22 to the third learning model 23, and converts the first learning model 21 into a machine based on the output generated data. A first process of learning is performed.

Next, in step S7, the first learning processing unit 11 determines whether or not learning is completed, that is, whether or not a predetermined learning end condition is satisfied. Return to step S6. Note that, as the learning termination condition here, for example, a condition provided for a loss during learning of the first learning model 21 can be considered.

In step S8, the processing unit 1 determines whether or not the second repetition end condition is satisfied. If Yes, the process ends, and if No, the process returns to step S1. As the second repetition end condition, for example, the number of repetitions (N2) can be considered.

Note that as the learning of the first learning model 21 progresses by repeating step S6, the feature amount calculated from the actual data/generated data by the first learning model 21 may change. Therefore, it is necessary to repeat the learning of the second learning model 22 (step S1) and the learning of the third learning model 23 (step S3), so the process returns to step S1. The model 23 is returned to the initial state before learning. By repeating the above, the performance of the first learning model 21 can be improved using the synthesized data (generated data).

As described above, according to the machine learning device 100 of the present embodiment, the third learning model 23 outputs the generated data in which the features of the real data are added to the synthetic data, and by using the generated data, The accurate first learning model 21 can be stably created. In other words, since the synthetic data is brought closer to the real data (realization processing) from the perspective of the first learning model 21 whose performance is to be improved using the synthetic data, a learning effect equivalent to the real data is expected. can.

Further, by repeating the entire process of FIG. 7, the accuracy of the first learning model 21 can be further improved.

Moreover, by using the above-described difference, it is possible to avoid a situation in which the third learning processing unit 13 generates generated data that is far from the synthesized data just to bring the synthesized data closer to the real data.

Also, by weighting according to the scene, such as changing the above-mentioned weights in the first half and the second half of the entire iterative process in FIG.

The number of feature values used in the first learning model 21 is as large as tens of thousands to billions. Therefore, it is impossible for humans to accurately grasp the characteristics of the feature amounts of the actual data and the synthetic data. According to the machine learning device 100 of the present embodiment, by generating generated data such that the first learning model 21 is deceived as real data, the highly accurate first learning model 21 is generated using such generated data. It can be created stably.

In addition, when learning the first learning model 21 using synthetic data (generated data), it is not necessary to provide a special mechanism to the first learning model 21 whose performance is to be improved. Learning model 21) can be effectively utilized.

In addition, by repeating the cycle of learning the third learning model 23 that performs realization processing to fill the gap between the real data and the synthetic data and learning the first learning model 21 whose performance is to be improved, dual performance improvement is achieved. can be planned. As a result, it can be expected that the performance of the first learning model 21 is significantly improved as compared with the conventional technology.

It should be noted that the program for executing the information processing in the above-described embodiment can be stored as a file in an installable format or an executable format on a CD-ROM, flexible disk (FD), CD-R, DVD (Digital Versatile Disk). ), USB (Universal Serial Bus) memory or other computer-readable recording medium. Also, the program may be provided or distributed via a network such as the Internet. Alternatively, the program may be provided by being incorporated in a ROM or the like in advance.

In addition, the program has a module configuration including each of the above functional configurations. Each functional unit is loaded onto the RAM, and each functional unit described above is generated on the RAM. Part or all of the functional units described above can also be realized using dedicated hardware such as ASIC (Application Specific Integrated Circuit) and FPGA (Field-Programmable Gate Array).

Although the embodiment has been described, the above embodiment is presented as an example and is not intended to limit the scope of the invention. The novel embodiments described above can be embodied in various other forms, and various omissions, replacements, and modifications can be made without departing from the scope of the invention. The above embodiments are included in the scope and gist of the invention, and are included in the scope of the invention described in the claims and equivalents thereof.

For example, the target data is not limited to image data such as photographed data/CG data, and may be sensor data (waveform data, etc.).

Further, when the second learning model 22 discriminates between the real data and the synthetic data (generated data), it is not limited to the direct discrimination method. Indirect identification methods such as AI for measuring differences and existing methods (Kullback-Leibler divergence, JS-divergence, etc.) may be employed.

Further, the calculation of the difference between the synthesized data and the generated data in the difference calculation processing unit 16 may be replaced with an AI trained in advance to measure the closeness between the actual data and the synthesized data.

In addition, when weighting the difference between the identification error and the synthetic data/generated data, a mechanism (reinforcement learning, etc.) that optimizes the weighting method itself is implemented in the entire cycle (repeating steps S1 to S8 in FIG. 7). may be incorporated into

Reference Signs List 1 processing unit 2 storage unit 11 first learning processing unit 12 second learning processing unit 13 third learning processing unit 14 estimation processing unit 15 setting unit 16 difference calculation processing Unit 17 Control unit 21 First learning model 22 Second learning model 23 Third learning model 100 Machine learning device 101 Processor 102 ROM 103 RAM 104 Input unit , 105... display section, 106... communication I/F, 107... HDD

Claims (5)

A first learning model for inputting data and outputting a feature amount of the data, and data input to the first learning model for inputting the feature amount output from the first learning model are real data. a second learning model that outputs a discrimination error as to whether or not the discrimination error output from the second learning model is input to output generated data in which the features of the real data are added to the synthetic data a storage unit that stores a third learning model;
Second learning for performing a second process of inputting the actual data, the synthetic data, and the generated data into the first learning model and machine-learning the second learning model based on the output feature amount. a processing unit;
performing a third process of inputting the feature amount output from the first learning model into the second learning model and performing machine learning of the third learning model based on the output identification error; a learning processing unit;
A first process of inputting the identification error output from the second learning model into the third learning model and performing machine learning of the first learning model based on the output generated data. a learning processing unit;
A machine learning device with The second process and the third process are repeated until the first repetition end condition is satisfied, and then the first process is performed. A series of processes is repeated until the second repetition end condition is met. The machine learning device of claim 1, iterating. further comprising a difference calculation processing unit that calculates a difference between the synthesized data and the generated data,
3. The machine learning device according to claim 1, wherein said third learning processing unit further uses said difference to perform said third processing. 4. The machine learning device according to claim 3, wherein said third learning processing unit weights said identification error and said difference and performs said third processing. A first learning model for inputting data and outputting a feature amount of the data, and data input to the first learning model for inputting the feature amount output from the first learning model are actual data. a second learning model that outputs a discrimination error as to whether or not the discrimination error output from the second learning model is input to output generated data in which the features of the real data are added to the synthetic data A machine learning method using a third learning model,
Second learning for performing a second process of inputting the actual data, the synthetic data, and the generated data into the first learning model and machine-learning the second learning model based on the output feature amount. a processing step;
performing a third process of inputting the feature amount output from the first learning model into the second learning model and performing machine learning of the third learning model based on the output identification error; a learning processing step;
A first process of inputting the identification error output from the second learning model into the third learning model and performing machine learning of the first learning model based on the output generated data. a learning processing step;
Machine learning methods, including

Images