JP7058202B2 - Information processing method and information processing system - Google Patents

Description

The present invention relates to an information processing system that executes a learning model and an information processing method performed by the information processing system.

In recent years, techniques related to reducing the weight of the learning model obtained by machine learning and applying it to an embedded system have been proposed. The technique proposed in this way includes a method for improving the performance of the learning model in the embedded system after application (see, for example, Patent Document 1 and Non-Patent Document 1).

International Publication No. 2017/038104

Benoit Jacob, 7 others, "Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference", [online], Internet <URL: https://arxiv.org/abs/1712.05877>

However, when the conversion content of the conversion tool that converts the learning model is unknown (that is, the conversion tool is a black box), the above-mentioned conventional technique cannot be used.

Therefore, the present invention provides an information processing method and the like that can improve the performance of the learning model after conversion even if the conversion content of the conversion tool for converting the learning model is unknown.

In the information processing method according to one aspect of the present invention, the output data of the first learning model is used as training data by using a computer, and the output data of the second learning model obtained by the conversion of the first learning model is correctly answered. The first output data corresponding to the first input data to the first training model is input to the third training model obtained by the training used as data to acquire the second output data, and the first input data is used. It is an information processing method that acquires the first correct answer data and relearns the first learning model by using the first difference data corresponding to the difference between the second output data and the first correct answer data.

Further, the information processing system according to one aspect of the present invention is an information processing system including a computer, in which the computer uses the output data of the first learning model as training data and is converted by the first learning model. The first output data corresponding to the first input data to the first learning model is input to the third learning model obtained by learning using the output data of the obtained second learning model as correct answer data, and the second output is performed. Data is acquired, the first correct answer data for the first input data is acquired, and the first difference data corresponding to the difference between the second output data and the first correct answer data is used to obtain the first learning model. Relearn.

It should be noted that this comprehensive or specific embodiment of the present invention may be realized by a recording medium such as a device, an integrated circuit, a computer program or a computer-readable CD-ROM, and the device, system, method, integrated circuit, etc. It may be realized by any combination of a computer program and a recording medium.

In the information processing method and the information processing system of the present invention, even if the conversion content of the conversion tool for converting the learning model is unknown, the performance of the learning model after conversion can be improved.

FIG. 1 is a diagram for explaining an outline of learning of a learning model and application to an embedded system. FIG. 2 is a diagram for explaining an outline of the mechanism for reflecting the loss of the output of the model in the embedded system in the learning model in the embodiment. FIG. 3 is a block diagram showing a functional configuration of an information processing system that realizes the relearning process in the embodiment. FIG. 4 is a sequence diagram showing the flow of data in the embodiment in chronological order. FIG. 5 is a diagram for explaining an outline of a mechanism for reflecting a loss of model output in an embedded system in a learning model in a modified example of the embodiment.

(Knowledge that became the basis of the present invention)
The present inventors have found that the following problems arise with respect to the application of the above-mentioned model obtained by machine learning (hereinafter, also referred to as a learning model) to an embedded system by reducing the weight.

Embedded systems are usually inferior to the environment in which the learning model is built in terms of computing speed, memory size, and the amount of power that can be supplied. For example, for learning, a numerical expression such as FP32 (32-bit floating point) is used for higher accuracy, and a computer that can process a numerical expression of such a size in a shorter time. Is used. On the other hand, in the embedded system, the scale of the processing circuit is limited depending on the device to which the embedded system is embedded. In addition, for embedded systems used in environments where the available power is limited, such as automobiles, it is necessary to consider power consumption when selecting the processing circuit.

The learning model built on the computer as described above is implemented in the embedded system after being converted for weight reduction so that it can be executed at the required speed even under these restrictions. FIG. 1 is a diagram for explaining an outline of construction of a learning model and its application to an embedded system. In this example, a neural network type learning model LM is constructed by the learning process in the learning unit. The learning unit has, for example, a functional configuration constructed by executing a predetermined program on a personal computer that operates under the supply of commercial power, and performs learning processing using the numerical representation of FP32. Then, the weights 1A, 2A, and 3A corresponding to the nodes, which are the parameters of this neural network, are converted into an integer representation with a smaller number of digits such as 16 bits or 8 bits by a quantization tool, that is, quantized. Are the weights 1a, 2a and 3a. Since the learning model lm including these parameters has a smaller execution load than the original learning model LM, it contributes to shortening the processing time on an embedded system equipped with a processor having a slower processing speed and a smaller memory than that of a personal computer. do.

However, in the learning model lm obtained by the quantization in this way, the inference output as a result for the input may not match the inference by the original learning model. The learning model lm is usually inferior to the original learning model LM in inference accuracy. That is, the loss, which is the difference from the correct answer data for the input, tends to be larger in the learning model lm. However, when the embedded system plays a recognition function in, for example, the driving support of the automobile mentioned above or automatic driving, if false detection or non-detection of the object occurs due to insufficient accuracy of inference. , Involved in safety issues.

There are existing methods for improving the inference accuracy of learning models. For example, in the learning unit of FIG. 1, re-learning is executed using a method called an error back-propagation method. To roughly explain the flow of this method, first, a loss C, which is the difference between the inference output data B and the correct answer data for the input data corresponding to the output data B, is obtained, and the loss C is used from the output layer. The weights of the neurons are adjusted so as to go back to the previous stage (see the arrow of the alternate long and short dash line in Fig. 1).

However, this method is only for reducing the loss of inference executed in the learning unit in which a numerical expression capable of highly accurate inference is used. Even if the learning model LM is retrained so that this loss is small, the difference between the output data b, which is the result data of inference in the embedded system and is different from the output data B, and the correct answer data is small. It is not always the case. In other words, even if the loss of the output data B is eliminated by this method, the difference of the output data b with respect to the correct answer data is not always eliminated. Such a difference that cannot be eliminated by the embedded system leads to erroneous detection or non-detection of an object outside the vehicle in the above-mentioned example of an automobile.

Further, Patent Document 1 includes a step of comparing the inference result in the CPU of the personal computer corresponding to the learning unit and the inference result in the embedded chip corresponding to the embedded system of FIG. It has been disclosed. The next step is to tune the code, but the disclosure does not include details on how to use the results of this comparison in tuning, thus improving the accuracy of inference in embedded systems. I don't know.

In the technique disclosed in Non-Patent Document 1, a conversion tool corresponding to the above-mentioned quantization tool is embedded in the inference path in the learning model in the personal computer corresponding to the above-mentioned learning unit. As a result, the learning model implemented in the embedded system is virtually constructed on the personal computer. Then, re-learning is performed using the inference result of this virtually constructed learning model. This solves the problem that the accuracy of inference in the embedded system cannot be improved due to the disagreement between the learning model on the personal computer and the learning model on the embedded system. Here, the method disclosed in Non-Patent Document 1 is premised on the fact that the contents of the conversion tool are clear. However, the contents of the conversion tool such as the quantization tool are not disclosed by the vendor who provides the framework including the tool, and it is generally a black box for the user. That is, the user of the conversion tool cannot use this method in order to eliminate the loss of the output data b of the learning model lm in the embedded system.

The present invention provides an information processing method that can be used to eliminate an error in the output of a learning model in an embedded system even for a user of a neural network who cannot know the contents of the conversion tool as described above. It is a thing.

In the information processing method according to one aspect of the present invention, the output data of the first learning model is used as training data by using a computer, and the output data of the second learning model obtained by the conversion of the first learning model is correctly answered. The first output data corresponding to the first input data to the first training model is input to the third training model obtained by the training used as data to acquire the second output data, and the first input data is used. It is an information processing method that acquires the first correct answer data and relearns the first learning model by using the first difference data corresponding to the difference between the second output data and the first correct answer data.

As a result, re-learning can be executed based on the difference (loss) of the output data that is the result of inference of the embedded system with respect to the correct answer data. Therefore, even if the conversion content of the conversion tool that converts the learning model is unknown, the conversion is performed. The performance of the later learning model can be improved. Specifically, the error of the output of the learning model in the embedded system can be eliminated, and the inference accuracy of the embedded system can be improved.

Further, for example, the difference data corresponding to the difference between the output data of the second learning model and the correct answer data for the input data to the first learning model is used as the training data, and the output data of the first learning model and the first training model are used. The first difference data is input to the fourth learning model obtained by learning using the difference data corresponding to the difference from the correct answer data corresponding to the input data of the first training model as the correct answer data, and the second difference data is acquired. Then, the first learning model may be retrained using the second difference data.

For example, if the difference in different environments is used as it is for re-learning, the weighting adjustment by the back-propagation method is excessive or insufficient, and as a result of re-learning, the inference accuracy does not improve or worsens. Can also occur. However, with this configuration, as the above-mentioned difference used for re-learning, a more appropriate one that is less likely to cause such a situation can be obtained, and the efficiency of re-learning can be improved.

Further, for example, the third learning model may be obtained by learning using the input data corresponding to the output data of the first learning model as further learning data, or the third learning model may be obtained by the first learning model. It may be obtained by learning using the learning parameters corresponding to the output data of the learning model as further learning data.

As a result, more appropriate output data of the embedded system to be inferred can be obtained, and the efficiency of re-learning can be improved.

Further, for example, the first learning model and the second learning model are neural network type learning models, and the learning parameters may be weights corresponding to the nodes of the neural network.

As a result, it is possible to suppress a decrease in accuracy of a high-precision neural network obtained in a learning environment, which is a concern when applied to an embedded system with severe restrictions on the execution environment.

Further, for example, the conversion of the first learning model is to reduce the weight of the neural network.

This improves the performance of the learning model, which is lighter because it is used in an embedded system that has stricter resource restrictions than the environment at the time of construction.

Further, for example, the output data of the first learning model may be used as training data, and the output data of the second learning model may be used as correct answer data to train the third learning model. Further, the difference data corresponding to the difference between the output data of the second learning model and the correct answer data for the input data to the first learning model is used as the training data, and the output data of the first training model and the first training model are used. The fourth learning model may be trained by using the difference data corresponding to the difference from the correct answer data corresponding to the input data of the training model as the correct answer data.

This simulates the loss of the output data of the second training model using data that can be used even under the condition that the content of the quantization tool for obtaining the training model used in the embedded system is a black box. Is obtained. This simulation feeds back to the first learning model to improve the performance of the learning model used in the embedded system.

Further, the information processing system according to one aspect of the present invention is an information processing system including a computer, in which the computer uses the output data of the first learning model as training data and is converted by the first learning model. The first output data corresponding to the first input data to the first learning model is input to the third learning model obtained by learning using the output data of the obtained second learning model as correct answer data, and the second output is performed. Data is acquired, the first correct answer data for the first input data is acquired, and the first difference data corresponding to the difference between the second output data and the first correct answer data is used to obtain the first learning model. Relearn.

As a result, re-learning can be executed based on the difference (loss) of the output data that is the result of inference of the embedded system with respect to the correct answer data. Therefore, even if the conversion content of the conversion tool that converts the learning model is unknown, the conversion is performed. The performance of the later learning model can be improved. Specifically, the error of the output of the learning model in the embedded system can be eliminated, and the inference accuracy of the embedded system can be improved.

Hereinafter, the information processing method and the information processing system according to the embodiment will be specifically described with reference to the drawings.

The following embodiments show a comprehensive or specific example of the present invention. Therefore, the numerical values, shapes, materials, components, arrangement and connection forms of the components, steps (processes), order of steps, etc. shown in the following embodiments are merely examples and do not limit the present invention. not. Then, among the components in the following embodiments, the components not described in the independent claims are components that can be arbitrarily added. Further, each figure is a schematic view and is not necessarily exactly shown.

(Embodiment)
FIG. 2 is a diagram for explaining an outline of the mechanism for reflecting the difference between the correct answer data and the output of the model in the embedded system in the learning model in the embodiment. Hereinafter, the situation will be described in comparison with the situation shown in FIG. 1 lacking this mechanism.

The neural network learning unit 10 is realized by executing a predetermined program by a processor in, for example, a personal computer or the like. The neural network learning unit 10 handles a numerical expression capable of high-precision inference, such as FP32, and constructs a learning model LM by learning. In addition, the neural network learning unit 10 also executes reasoning on the input data by the constructed learning model LM, and relearns the learning model LM using the difference between the correct answer data and the output data which is the result of the reasoning. Can be done. These points are common to those shown in FIG. The learning model LM is an example of the first learning model in the present embodiment.

The quantization tool 15 is common to that shown in FIG. 1, and obtains a lighter learning model lm by quantizing the weight of the learning model LM constructed in the neural network learning unit 10. Quantization tools are included in vendor-provided deep neural network software frameworks such as TensorRT, XNNC (Xtensa Neural Network Compiler), CDNN (CEVA Deep Neural Network), and are a black box for users.

The learning model lm is also the same as that shown in FIG. 1, and is an embedded system 20 that handles numerical expressions such as int16 (integer 16 bits) or int8 (integer 8 bits) that have a smaller processing load than the neural network learning unit 10. Is implemented in. The learning model lm is an example of the second learning model in the present embodiment.

The feedback unit 30 is an element not included in the mechanism of FIG. 1 that feeds back the loss of the output data of the learning model lm to the learning model LM.

The first conversion model CM1 included in the feedback unit 30 is a learning model obtained by learning using the output data which is the result of the inference of the learning model LM as the learning data and the output data of the learning model lm as the correct answer data. Yes, for example by a neural network. The function of the first conversion model CM1 in the feedback unit 30 is to receive the output data B corresponding to a certain input data to the learning model LM as an input and acquire the output data bb simulating the result of inference by the learning model lm. be. The fact that the axis of the arrow indicating the output from the learning model lm is a broken line indicates that the output data bb is simulated and is not actually output by the learning model lm. The first conversion model CM1 is an example of the third learning model in the present embodiment. Further, the output data B is an example of the first output data in the present embodiment, and the input data corresponding to the first output data is an example of the first input data in the present embodiment. Further, the output data bb is an example of the second output data in the present embodiment. That is, the second output data is a simulation of the output data of the second learning model.

In the feedback unit 30, a loss calculation unit (not shown in FIG. 2, which will be described later) further acquires correct answer data for the first input data, and calculates a loss c which is a difference between the correct answer data and the output data bb. The correct answer data for the first input data, which is compared with the output data bb by the loss calculation unit, is an example of the first correct answer data in the present embodiment. Further, the loss c is an example of the first difference data in the present embodiment.

The second conversion model CM2 uses the loss, which is the difference between the output data corresponding to the input data, which is the result of the inference of the training model lm, and the correct answer data for the input data, as the training data, and also uses the loss of the training model LM. It is a learning model obtained by learning using the difference between the output data and the correct answer data for the input data corresponding to the output data as the correct answer data, and is, for example, by a neural network. The function of the second conversion model CM2 in the feedback unit 30 is to receive the loss c as an input and acquire the loss CC simulating the loss of the output data from the learning model LM. The second conversion model CM2 is an example of the fourth learning model in the present embodiment, and the loss CC is an example of the second difference data in the present embodiment. That is, the second difference data simulates the loss of the output data of the learning model LM.

The neural network learning unit 10 uses the loss CC output by the second conversion model CM2 to readjust, that is, relearn the weights of the learning model LM.

By such a mechanism, first, simulated output data of the learning model LM on the embedded system 20 is generated from the output data of the learning model LM on the personal computer. Then, the learning model LM is relearned using the difference between the simulated output data and the correct answer data for the corresponding input data. As a result, feedback of the result of inference by the embedded system to the learning model LM can be appropriately performed even when the quantization tool 15 is a black box and the content of the conversion by the quantization tool 15 is unknown. ..

The loss c may be used for re-learning of the learning model LM without being converted into the loss CC by the second conversion model CM2. However, in the re-learning using the loss c as it is due to the difference in specifications between the learning model LM and the learning model lm in which the inference result is simulated, for example, the difference in the numerical expression used, the learning model LM itself. It is possible that the accuracy of the inference of is deteriorated. In such a case, the accuracy of the learning model lm obtained by converting the learning model LM after re-learning may not be improved. The second conversion model CM2 is included in the feedback unit 30 as necessary for the purpose of avoiding such a situation.

Further, the re-learning process in the present embodiment does not replace the re-learning process shown in FIG. Although not shown in FIG. 2, the re-learning shown in FIG. 1 may also improve the accuracy of the learning model LM.

Next, a configuration that realizes the re-learning process in the present embodiment will be described. FIG. 3 is a block diagram showing a functional configuration of the information processing system 100 that realizes the above-mentioned re-learning process in the present embodiment.

The information processing system 100 is a system that executes learning and relearning of a learning model, and is composed of one or a plurality of computers. The learning and re-learning of this learning model is performed by the learning unit 10 of the neural network (denoted as NN using the acronym of the English notation Natural Network in FIG. 3) also shown in FIG. In addition to the neural network learning unit 10, the information processing system 100 includes a first conversion unit 31, a loss calculation unit 33, correct answer data 35, and a second conversion unit 37 as functional components.

The first conversion unit 31 is a functional component realized by constructing the first conversion model CM1 of FIG. 2 by a processor included in the computer constituting the information processing system 100. The first conversion unit 31 converts the output data B showing the inference result of the learning model LM received as an input by the constructed first conversion model CM1 and simulates the output data showing the inference result of the learning model lm. The output data bb is acquired and output.

The loss calculation unit 33 is a functional component provided by a predetermined program executed by a processor included in the computer constituting the information processing system 100. The loss calculation unit 33 receives the output data bb and the first correct answer data as inputs, calculates the difference between them, and outputs this difference as the loss c.

The correct answer data 35 is data held in a storage device included in a computer constituting the information processing system 100, and the first correct answer data is included in the correct answer data 35. Further, the correct answer data 35 for obtaining the loss used for the re-learning by the neural network learning unit 10 shown in FIG. 1 is also included in the correct answer data 35.

The second conversion unit 37 is a functional component realized by constructing the second conversion model CM2 of FIG. 2 by the processor included in the computer constituting the information processing system 100. The second conversion unit 37 converts the loss c of the output data of the learning model lm received as an input by the second conversion model CM2 constructed, acquires the loss CC, and outputs the loss CC. The loss CC is used by the neural network learning unit 10 for re-learning the learning model LM.

The first conversion unit 31, the loss calculation unit 33, the correct answer data 35, and the second conversion unit 37 may be realized on a computer that realizes the neural network learning unit 10, or may be realized on another computer. May be done. Further, the neural network learning unit 10 itself may be realized on one computer or may be realized on a plurality of computers.

Next, a procedure of data flow and processing in the above mechanism including the components described with reference to FIG. 3 will be described.

FIG. 4 is a sequence diagram showing the flow of data in the present embodiment in chronological order. In the following description, the data flow will be described by dividing it into the following four phases.

1st phase: From construction of learning model to implementation in embedded system 2nd phase: Construction of 1st transformation model by learning 3rd phase: Construction of 2nd transformation model by learning 4th phase: Re-learning

In FIG. 4, the correct answer data 35 is not shown for the sake of readability.

First, in the first phase, the neural network learning unit 10 constructs a learning model LM by learning using learning data and correct answer data (step S10). Although not shown, the accuracy of inference of the learning model LM may be improved by re-learning as shown in FIG. 1 at this stage as well. The constructed learning model LM is input to the quantization tool 15 (step S11) and quantized (step S12). The quantized learning model lm is implemented in the embedded system 20 (step S13). As a result, the embedded system 20 is in a state where inference by the learning model lm can be executed.

In the second phase, the neural network learning unit 10 performs inference by the learning model LM constructed in the first phase (step S20). Further, in the embedded system 20, inference is performed by the learning model lm implemented in the first phase (step S22). These inferences are performed using the same input data. The input data may be processed in the embedded system 20.

The first conversion unit 31 which acquired the result of the inference by the learning model LM and the result of the inference by the learning model lm (steps S21 and S23) uses the result of the inference by the learning model LM as learning data and infersed by the learning model lm. The first conversion model CM1 is constructed by learning using the result of the above as correct answer data (step S24). The first conversion model CM1 constructed by the learning process performed in advance may be acquired from a memory or the like.

In the third phase, the loss calculation unit 33 acquires the result of the inference already made by the learning model lm (step S30), and also obtains the correct answer from the correct answer data 35 to the input data corresponding to the result of this inference (output data). Acquire data (step S31). Then, the loss calculation unit 33 calculates the loss (difference) from the correct answer data with respect to the result of inference by the learning model lm (step S32).

On the other hand, the neural network learning unit 10 acquires the correct answer data for the input data corresponding to the inference result (output data) already performed by the learning model LM from the correct answer data 35 (step S34), and the inference result by the learning model LM. The loss (difference) from the correct answer data for is calculated (step S35).

The second conversion unit 37 that has acquired the loss of the inference result of the learning model lm calculated by the loss calculation unit 33 and the loss of the inference result of the learning model LM calculated by the neural network learning unit 10 (steps S33 and S36) is learning. The second conversion model CM2 is constructed by learning (step S37) in which the loss of the inference result of the model lm is used as learning data and the loss of the inference result of the learning model LM is used as correct answer data. The second conversion model CM2 constructed by the learning process performed in advance may be acquired from a memory or the like.

From the first phase to the third phase, the feedback unit 30 shown in FIG. 2 is constructed in the information processing system 100. The fourth phase is executed by the information processing system 100 that has reached this state.

In the fourth phase, the result of inference (output data B) executed by the learning model LM in the neural network learning unit 10 is converted by the first conversion model CM1 in the first conversion unit 31 (steps S40 to S42). .. The converted inference result (output data bb) is a simulation of the result inferred by the learning model lm from the input data corresponding to the inference result by the learning model LM (output data B).

The loss calculation unit 33 that has acquired the output data bb (step S43) acquires the correct answer data (first correct answer data) for the input data corresponding to the result of inference by the learning model LM (output data B) from the correct answer data 35 (step S43). Step S44). Then, the loss calculation unit 33 calculates the loss (difference) from the first correct answer data with respect to the result of inference by the simulated learning model lm (step S45).

The second conversion unit 37 converts the loss by the second conversion model CM2 (steps S46 and S47). The neural network learning unit 10 acquires this converted loss (step S48) and uses it for re-learning of the learning model LM (step S49).

When the weights 1A to 3A are updated by the re-learning of the learning model LM, the re-learning may be repeated by returning to the first phase and executing the processing after the quantization (step S12).

(Modification example)
Although the information processing method and the information processing system according to one or more embodiments have been described above based on the embodiments, the present invention is not limited to these embodiments. As long as it does not deviate from the gist of the present invention, various modifications that can be considered by those skilled in the art to the present embodiment may be included within the scope of one or more embodiments.

FIG. 5 is a diagram for explaining an example of such a modification. Hereinafter, the differences from the mechanism shown in FIG. 2 will be mainly described.

In the mechanism shown in FIG. 5 for reflecting the difference to the correct answer data in this modification in the learning model, the input to the first conversion model corresponds to the output data B in addition to the output data B of the learning model LM. The mechanism is different from the mechanism shown in FIG. 2 in that the input data to be input and the weights 1A to 3A which are the parameters of the learning model LM are included.

That is, in this modification, in the second phase, the output data B, the input data corresponding to the output data B, and the parameters of the learning model LM are used as the learning data for constructing the first conversion model CM1. ing. As described above, the parameters of the learning model LM, which is also used as the learning data for the learning of the first conversion model CM1, are examples of the learning parameters in this modification.

As described above, the first conversion model CM1 obtained by learning using more learning data can output the output data bb that simulates the output data of the learning model lm with higher accuracy.

The learning data input to the first conversion model CM1 and the data added to the output data B as input data for inference are only one of the input data corresponding to the output data B and the parameters of the learning model LM. There may be.

As another modification, the output data b may be used instead of the output data bb to calculate the loss c. That is, the output data based on the inference result of the learning model lm may be used for calculating the loss c without using the first conversion model CM1.

As another modification, a predicted change in the performance of the learning model lm may be presented depending on the presence or absence of conversion using the first conversion model CM1 and the second conversion model CM2. For example, when only the first conversion model CM1 is used, the loss calculation unit 33 estimates a change in the accuracy of the learning model lm from the difference between the calculated loss c and the loss C. Further, when the first conversion model CM1 and the second conversion model CM2 are used, the loss calculation unit 33 estimates the change in the accuracy of the learning model lm from the difference between the calculated loss CC and the loss C. Then, the presentation device separately provided in the information processing system presents the change in the accuracy of the estimated learning model lm. The presenting device may be a display, a projector, a speaker, or the like.

The present invention can be used in fields where the learning model is applied to a more restrictive execution environment than the environment at the time of construction, such as automobiles (including self-driving cars), household electrical equipment, wearable information terminals, and industrial equipment. , And other IoT-incorporated industrial fields.

1A, 1a, 2A, 2a, 3A, 3a parameters (weights)
10 Neural network learning unit 15 Quantization tool 20 Embedded system 30 Feedback system 31 First conversion unit 33 Loss calculation unit 35 Correct answer data 37 Second conversion unit 100 Information processing system B, b, bb Output data C, CC, c Loss CM1 1st conversion model CM2 2nd conversion model LM, lm learning model

Claims (9)

Using a computer,
The first learning model obtained by learning using the output data of the first learning model as training data and using the output data of the second learning model obtained by the conversion of the first learning model as correct answer data. The first output data corresponding to the first input data to the training model is input to acquire the second output data.
Acquire the first correct answer data for the first input data,
The first learning model is retrained using the first difference data corresponding to the difference between the second output data and the first correct answer data.
Information processing method. Difference data corresponding to the difference between the output data of the second learning model and the correct answer data for the input data to the first learning model is used as training data, and the output data of the first training model and the first training model are used. The first difference data is input to the fourth learning model obtained by learning using the difference data corresponding to the difference from the correct answer data corresponding to the input data of the above as the correct answer data, and the second difference data is acquired.
The first learning model is retrained using the second difference data.
The information processing method according to claim 1. The third learning model is obtained by learning using the input data corresponding to the output data of the first learning model as further learning data.
The information processing method according to claim 1 or 2. The third learning model is obtained by learning using the learning parameters corresponding to the output data of the first learning model as further learning data.
The information processing method according to any one of claims 1 to 3. The first learning model and the second learning model are neural network type learning models.
The learning parameter is a weight corresponding to a node of the neural network.
The information processing method according to claim 4. The transformation of the first training model is to reduce the weight of the neural network.
The information processing method according to claim 5. The output data of the first learning model is used as training data, and the output data of the second learning model is used as correct answer data to train the third learning model.
The information processing method according to any one of claims 1 to 6. Difference data corresponding to the difference between the output data of the second learning model and the correct answer data for the input data to the first learning model is used as training data, and the output data of the first training model and the first training model are used. The fourth learning model is trained using the difference data corresponding to the difference from the correct answer data corresponding to the input data of the above as the correct answer data.
The information processing method according to claim 2. An information processing system equipped with a computer
The computer
The first learning model obtained by learning using the output data of the first learning model as training data and using the output data of the second learning model obtained by the conversion of the first learning model as correct answer data. The first output data corresponding to the first input data to the training model is input to acquire the second output data.
Acquire the first correct answer data for the first input data,
The first learning model is retrained using the first difference data corresponding to the difference between the second output data and the first correct answer data.
Information processing system.

Images