JP2024070372A - Learning method, information processing system, program and learning model - Google Patents

Abstract

[Problem] To provide a learning method, information processing system, program, and learning model capable of learning using data for which features according to rules or the like are not explicitly given. [Solution] According to one aspect of the present invention, a learning method executed by an information processing system is provided. This learning method includes an output step, a reception step, and a learning step. In the output step, the learning model outputs data generated during the learning process, either in response to given conditions or without given conditions. In the reception step, input of a judgment result of whether the data is true or false by a judgment means is accepted. In the learning step, learning of the learning model is progressed based on the judgment result. [Selected Figure] Figure 2

Description

The present invention relates to a learning method, an information processing system, a program, and a learning model.

In deep learning and other machine learning methods, a learning model is trained using given data. At this time, the data has some characteristics. For example, data with characteristics means that data on the motion of a mass point follows Newton's laws of motion, or that heat transport follows the laws of thermodynamics. In addition, in a task such as recognizing a person's age from their face, it can be assumed that face data is governed by folkloric or social scientific laws. It has also been proposed to generate learning data to improve the accuracy of a learning model (see, for example, Patent Document 1).

JP 2021-96511 A

By the way, various data can be used for learning, such as experimental values, measured values, mechanically generated values, etc., but this data may not contain features that follow laws, etc. In particular, mechanically generated data does not have explicit features that follow laws, etc., so there is no choice but to expect the model to read those features from the data during the learning process.

In consideration of the above circumstances, the present invention provides a learning method, information processing system, program, and learning model that can learn using data for which characteristics according to rules, etc. are not explicitly given.

According to one aspect of the present invention, a learning method executed by an information processing system is provided. This learning method includes an output step, a reception step, and a learning step. In the output step, the learning model outputs data generated during the learning process, depending on given conditions or without given conditions. In the reception step, input of the judgment result of the judgment means as to whether the data is true or false is accepted. In the learning step, learning of the learning model is progressed based on the judgment result.

According to one aspect of the present invention, it is possible to perform learning without preparing sufficient data with explicit features.

1 is a diagram showing a configuration of an information processing device 1. 1 is a block diagram showing a functional configuration of an information processing device 1. FIG. FIG. 2 illustrates an example of the configuration of a learning unit 112. FIG. 2 is a diagram illustrating an example of the configuration of a learning unit 112 when performing pre-learning. 11A and 11B are diagrams showing examples of adjustment of a determination unit 121. 11 is a diagram showing an example of a learning result by the information processing device 1. FIG. FIG. 13 is a diagram showing an example of a learning result obtained by another conventional method.

Embodiments of the present disclosure will be described below with reference to the drawings. The various features shown in the embodiments below can be combined with each other.

The program for implementing the software used in this embodiment may be provided as a non-transitory computer-readable recording medium, or may be provided so that it can be downloaded from an external server, or may be provided so that the program is started on an external computer and its functions are implemented on a client terminal (so-called cloud computing).

In this embodiment, a "unit" may also include, for example, a combination of hardware resources implemented by a circuit in the broad sense and software information processing that can be specifically realized by these hardware resources. In addition, this embodiment handles various types of information, which may be represented, for example, by physical values of signal values representing voltage and current, high and low signal values as a binary bit collection consisting of 0 or 1, or quantum superposition (so-called quantum bits), and communication and calculations may be performed on a circuit in the broad sense.

In the broad sense, a circuit is a circuit realized by at least appropriately combining a circuit, circuitry, a processor, and memory. In other words, it includes application specific integrated circuits (ASICs), programmable logic devices (e.g., simple programmable logic devices (SPLDs), complex programmable logic devices (CPLDs), and field programmable gate arrays (FPGAs)), etc.

1 is a diagram showing the configuration of an information processing device 1. As shown in the figure, the information processing device 1 has a processing unit 11, a storage unit 12, a temporary storage unit 13, an external device connection unit 14, and a communication unit 15, and these components are electrically connected via a communication bus 16 inside the information processing device 1.

The processing unit 11 is realized, for example, by a central processing unit (CPU) and operates according to a predetermined program stored in the memory unit 12 to realize various functions.

The storage unit 12 is a non-volatile storage medium that stores various information. This is realized by a storage device such as a hard disk drive (HDD) or a solid state drive (SSD). The storage unit 12 can also be arranged in a separate device that can communicate with the information processing device 1.

The temporary storage unit 13 is a volatile storage medium. It is realized by a memory such as a random access memory (RAM), and stores information (arguments, arrays, etc.) that is temporarily required when the processing unit 11 operates.

The external device connection unit 14 is a connection unit that conforms to standards such as Universal Serial Bus (USB) and High-Definition Multimedia Interface (HDMI (registered trademark)), and allows the connection of input devices such as keyboards and display devices such as monitors.

The communication unit 15 is, for example, a communication means conforming to the Local Area Network (LAN) standard, and realizes communication between the information processing device 1 and the local area network or a network such as the Internet via the local area network.

In addition, the information processing device 1 can be a general-purpose server computer, a personal computer, etc., and it is also possible to configure the information processing device 1 using multiple computers.

2. Functions of the information processing device 1 Next, the functions of the information processing device 1 will be described. The information processing device 1 constitutes an information processing system, and operates according to a program to realize each functional unit described later. This program causes at least one computer to execute each step of a learning method described later. Specifically, the processing unit 11, which is hardware, operates based on a program stored in the storage unit 12, i.e., software, to realize each functional unit described later. At this time, the processing unit 11 operates the storage unit 12, the temporary storage unit 13, the external device connection unit 14, and the communication unit 15 as necessary.

The information processing system described here is an information processing system consisting of at least one device. Here, an information processing system including an information processing device 1 as the at least one device will be described as an example. This information processing system includes at least one processor capable of executing programs to operate each unit described below. This processor corresponds to, for example, the processing unit 11.

The learning model realized by the information processing system is learned using a learning method described below, and the results of the learning are stored in the memory unit 12. This stored learning model can be used as a trained learning model in other information processing devices such as computers.

Figure 2 is a block diagram showing the functional configuration of the information processing device 1. As shown in the figure, the information processing device 1 realizes the functions of a learning device 110 and a judgment device 120 through a program. Note that in Figure 2, the learning device 110 and the judgment device 120 are realized in the information processing device 1, but the judgment device 120 may be realized in a separate information processing device that can communicate via a network or the like, or may be realized as a device having a configuration separate from the information processing device.

The learning device 110 includes a condition receiving unit 111, a learning unit 112, an output unit 113, and a receiving unit 114. The judgment device 120 includes, for example, a judgment unit 121. The condition receiving unit 111 receives conditions for learning. For example, when learning to predict the trajectory of a thrown mass point, the initial velocity and the angle at the time of throwing are the conditions, when learning the design of a two-dimensional laminar flow wing (such as the main wing of a model airplane), the lift coefficient is the condition, and when learning the design of a turbine wing (such as the wing of the rotating part of an aircraft engine), the outflow angle and the loss coefficient are the conditions. This condition may be one numerical value, may be specified as multiple numerical values or a numerical range, or may not be a condition at all. The learning unit 112, the output unit 113, and the receiving unit 114 respectively execute a learning step, an output step, and a receiving step, which will be described later.

The judgment unit 121, which is a judgment means, judges whether the data output by the output unit 113 in the output step described later is true or false, based on the same conditions as those accepted by the condition acceptance unit 111. The judgment device 120 including the judgment unit 121 can be realized by software other than the learning device 110, such as analysis software, design software, or simulation software, and preferably judges the true or false of the data output in the output step based on the calculation results of the other software capable of calculation based on physical laws, for example, judging whether the data output in the output step is physically correct or not, and judging it as "true" if it is correct, and "false" if it is not correct. Specifically, it judges "true" if the error between the value calculated based on the data output in the output step and the value calculated based on physical laws is smaller than a predetermined value, and judges "false" if the error is equal to or greater than a predetermined value. The determination device 120 does not need to perform physical verification. For example, when a mass point is thrown, the trajectory does not rise and then fall, and does not rise again, so the trajectory does not become a so-called jagged trajectory. Software that analyzes shapes can be used to detect the presence or absence of jagged edges. In addition, when performing a sensory evaluation, for example when learning about perfume blending, the results can be judged by a person such as an appraiser, and the judgment can be performed by a person.

The learning device 110 may also be provided with a pre-judgment unit (not shown) so that if the generated data is clearly abnormal, the data is not judged by the judgment unit 121 and the data is generated again. Examples of cases where the generated data is clearly abnormal include cases where it contradicts a simple rule, where the upper and lower limits of the generated value are exceeded, where the relationship between the output value and the input value (such as the upper and lower limits of a ratio) is broken, etc.

3. Overview of Learning Next, an overview of the learning performed by the learning device 110 will be described. The learning method described here is a learning method executed by an information processing system (information processing device 1). This learning method includes an output step, a reception step, and a learning step.

In the output step, the output unit 113 outputs to the determination unit 121 data that the learning model included in the learning unit 112 has generated during the learning process according to given conditions, i.e., conditions accepted by the condition acceptance unit 111, or without giving any conditions. At this time, the output unit 113 converts the data into a form that can be input by the determination unit 121. The determination unit 121 determines whether this data is true or false with respect to the conditions.

In the reception step, the reception unit 114 receives an input of the result of the judgment made by the judgment unit 121 as to whether the data is true or false. In the learning step, the reception unit 114 proceeds with learning of the learning model included in the learning unit 112 based on the judgment result received. In this way, the learning device 110 performs learning using the result of the judgment made by the judgment device 120. The judgment of true or false is, for example, a judgment of true or false with respect to a condition of the data.

4. Specific Example of Learning Fig. 3 is a diagram showing a configuration example of the learning unit 112. The learning unit 112 includes a generator 1121 and a discriminator 1122, which are learning models, and also includes a comparison unit 1123. Note that in Fig. 3, the comparison unit 1123, the condition reception unit 111, the output unit 113, and the reception unit 114 are omitted.

The generator 1121 is a type of neural network also called a generator, and generates data based on the conditions and random noise received by the condition receiving unit 111. The discriminator 1122 is a type of neural network also called a discriminator, and discriminates whether the data generated by the generator 1121 is true or false. The comparison unit 1123 executes a comparison step, in which the judgment result by the judgment unit 121 is compared with the discrimination result by the discriminator 1122.

In the learning step, the generator 1121 and the discriminator 1122 are trained based on the results of the comparison in the comparison step. In this learning, the generator 1121 is trained to generate data that the discriminator 1122 discriminates as true, and the discriminator 1122 is trained to discriminate the truth of the data in the same way as the result of the judgment by the judgment unit 121. This learning is a learning method similar to generative adversarial networks (GANs). In a generative adversarial network, the generator performs learning with the aim of causing the discriminator to make a mistake in discrimination, that is, to discriminate the data generated by the generator from the teacher data, and the discriminator performs learning with the aim of discriminating the teacher data as "true" and the data generated by the generator as "false". In response to this, the generator 1121 performs learning in the same way as the generator of the generative adversarial network, and the discriminator 1122 performs learning in the same way as in the case where data generated by the generator 1121 that is determined to be "true" by the determination unit 121 is regarded as training data for the generative adversarial network. Therefore, the specific methods for training the generator 1121 and the discriminator 1122 are similar to those for training in the generative adversarial network.

By having the generator 1121 and the discriminator 1122 perform this type of learning, as the learning progresses, the accuracy of the data generated by the generator 1121 increases, and the likelihood that the judgment unit 121 will generate data that is judged to be "false" decreases, meaning that it becomes possible to generate data that does not contradict physical laws, etc.

5. Pre-Learning Meanwhile, in the initial state, that is, in a state where no learning has been performed, the generator 1121 has a low probability of generating data that the determination unit 121 determines to be “true”, and the discriminator 1122 also has a low probability of making a classification similar to the determination result of the determination unit 121. Therefore, the time required for learning by the learning unit 112 may be long.

In order to avoid the time required for this learning from increasing, in the learning step, the generator 1121 may be pre-trained using teacher data before learning based on the results of the comparison in the comparison step. Since the generator 1121 can also operate as a neural network, in the pre-learning, the generator 1121 is operated as a general neural network and learning is performed using teacher data. In addition, in the learning step, the generator 1121, or both the generator 1121 and the discriminator 1122 may be pre-trained using teacher data before learning based on the results of the comparison in the comparison step. In this pre-learning, the generator 1121 is trained to generate data that the discriminator 1122 discriminates as true, and the discriminator 1122 is trained to discriminate the teacher data as true. Note that when only the generator 1121 is pre-trained, another discriminator is used instead of the discriminator 1122. In other words, pre-learning is equivalent to operating the generator 1121 and the discriminator 1122 as a generative adversarial network. In this pre-learning, the generator 1121 generates more than twice the predetermined number of data, for example, much more data than the predetermined number, and from this generated data, a predetermined number of data with low error and a predetermined number of data with high error can be used. This error is calculated by the determination unit 121 and is the difference between a value calculated based on data and a value based on physical laws. For example, if the number of training data is desired to be 64, the generator 1121 is made to generate about 10 times that amount of data, and from this generated data, 32 data with low error and 32 data with high error are used as training data.

The teacher data used in pre-learning may be data generated by the generator 1121 and determined to be "true" by the determination unit 121. FIG. 4 is a diagram showing an example of the configuration of the learning unit 112 when performing pre-learning. As shown in the figure, the learning unit 112 includes a providing unit 1124. The providing unit 1124 accumulates data determined to be "true" by the determination unit 121 and provides it to the discriminator 1122 as appropriate. "As appropriate" means that in pre-learning, the data generated by the generator 1121 and the teacher data are input randomly to the discriminator 1122, and therefore are not always provided.

In addition, in pre-learning, when the conditions are given as a range of numerical values, pre-learning may be performed with a predetermined number of numerical values included in the range of numerical values as the conditions. For example, when learning is performed with the condition that the lift coefficient is between 0 and 1.6, learning is performed so that any value between 0 and 1.6 can ultimately be input as the condition. However, in pre-learning, for example, learning can be performed with four conditions, 0.1, 1.5, 0.95, and 1.58, to improve the learning speed of the learning unit 112.

6. Adjustment of the Determination Means If the determination unit 121, which is the determination means, is adjustable, the speed of learning of the learning unit 112 can be increased by adjusting it. For example, if the determination unit 121 can set an upper limit value and a lower limit value for the reference value, the setting is performed as shown in FIG. 5. FIG. 5 is a diagram showing an example of adjustment of the determination unit 121. The reference value shown by a solid line in FIG. 5 indicates that the condition and the value based on the data generated by the generator 1121 according to the condition match. This reference value, for example, determines that the lift coefficient, which is the condition of the two-dimensional laminar wing to be designed, matches the lift coefficient of the two-dimensional laminar wing created according to the data (shape) of the two-dimensional laminar wing generated by the generator 1121 according to this condition, as "true." In the example shown in FIG. 5, when the condition is 0.8, the value based on the data generated by the generator 1121 is 0.8 only when it is 0.8. When the judgment unit 121 is adjustable, the judgment unit 121 judges the data to be "true" when the data is equal to or less than the upper limit of the reference value and equal to or greater than the lower limit of the reference value. For example, when the condition is 0.8 and the adjustable ranges of the upper limit and lower limit are both 0.2, if the value based on the data generated by the generator 1121 is equal to or less than 1.0 and equal to or greater than 0.6, the data is judged to be "true". It is desirable that the upper limit is lowered as the learning of the learning model progresses, and the lower limit is raised as the learning of the learning model progresses. In addition, the upper limit and lower limit may take different values depending on the value of the condition.

If the judgment unit 121, which is the judgment means, is capable of handling this, the judgment unit 121 may include the difference between the data and the reference value in the judgment result. In this case, the learning unit 112 will perform learning based on the difference between the data and the reference value.

6. Formulation In the above-described configuration, generated data is verified by a physical model without using teacher data, and if the error is large, it is judged as "false", and if the error is small, it is judged as "true", the classifier 1122 learns so as to be able to make the same judgment as the physical model, and the generator 1121 learns so as to deceive the classifier 1122, that is, to generate data that the classifier 1122 judges as "true". Note that separate software or the like can be used as the physical model.

Here, these will be explained using mathematical expressions. In the above-mentioned configuration, the objective is to generate x that satisfies mathematical expression 1. In other words, if x satisfies mathematical expression 1, it is determined to be "true," and if x does not satisfy mathematical expression 1, it is determined to be "false," and learning is performed. In this case, the objective function can be expressed by mathematical expression 3 using the set shown in mathematical expression 2.

However, since a set that satisfies Equation 2 is often 0, Equation 4 is also 0. For this reason, the set shown in Equation 5 is defined by Equation 6, where Equation 7 is used. Furthermore, the measure of Equation 5 is not 0, and an expected value can be calculated. Therefore, the objective function is as shown in Equation 8. Furthermore, in optimizing a neural network, optimization shown in Equation 9 is performed, where h(i) is a function of the epoch number i, and is a function that converges to 0 as i becomes larger.

7. Example of Use Next, an example of use of the above-mentioned configuration will be described. Suppose there is data shown in Expression 10. Assume that x of this data satisfies Expression 1 for a certain function f and parameter θ. In this case, the objective is to find x that satisfies Expression 12 for a given value shown in Expression 11.

Here, as a specific example, the case of wing design will be described. When the wing shape is expressed as a vector of a series of points, it can be expressed by Equation 13. Furthermore, for a certain x, the lift coefficient θ at a determined Reynolds number and angle of attack is uniquely determined. If function g is a function (shown in Equation 14) that calculates θ from x, Equation 1 can be expressed by Equation 15 using function g. In this case, searching for x that satisfies Equation 12 for a given value shown in Equation 11 is equivalent to searching for a shape x that gives a given lift coefficient (Equation 11).

8. Example of Results Next, an example of trial results by the above-mentioned configuration is shown. FIG. 6 is a diagram showing an example of learning results by the information processing device 1, and FIG. 7 is a diagram showing an example of learning results by another conventional method. In the examples shown in FIG. 6 and FIG. 7, the conditions are specified from 0 to 1.58, in increments of 1.58/1000. As is clear from FIG. 6 and FIG. 7, the learning by the information processing device 1 was able to obtain better results than the conventional method over the entire range of conditions.

9. Others The learning method described above is a learning method executed by an information processing system, and includes an output step, a receiving step, and a learning step.

The information processing device 1 may be in an on-premise form or in a cloud form. As an information processing device 1 in a cloud form, for example, the above functions and processes may be provided in the form of SaaS (Software as a Service) or cloud computing.

The present invention may be provided in the following aspects:

(1) A learning method executed by an information processing system, comprising an output step, a reception step, and a learning step, in which the output step outputs data generated by a learning model during the learning process in accordance with given conditions or without the conditions being given, the reception step receives input of a judgment result of the truth or falsity of the data by a judgment means, and the learning step progresses the learning of the learning model based on the judgment result.

(2) In the learning method described in (1) above, the determination means determines whether the data is true or false based on physical laws.

(3) In the learning method described in (2) above, the determination means determines whether the data corresponds to the condition.

(4) A learning method according to any one of (1) to (3) above, further comprising a comparison step, the learning model comprising a generator and a classifier, the generator generating the data, the classifier classifying the authenticity of the data, the comparison step comparing the judgment result with the classification result by the classifier, the learning step progressing the learning of the generator and the classifier based on the comparison result by the comparison step, the learning training training the generator to generate data that the classifier classifies as true, and the classifier to classify the authenticity of the data in the same way as the judgment result by the judgment means.

(5) In the learning method described in (4) above, in the learning step, pre-learning of the generator is carried out using teacher data prior to learning based on the results of the comparison in the comparison step, and the pre-learning causes the generator to operate as a neural network using the teacher data.

(6) In the learning method described in (4) above, in the learning step, prior to learning based on the results of the comparison in the comparison step, pre-learning of the generator and the discriminator is carried out using training data, and the pre-learning is a learning method in which the generator is trained to generate data that the discriminator will identify as true, and the discriminator is trained to identify the training data as true.

(7) A learning method according to any one of (4) to (6) above, in which the determination means calculates the error between a value calculated based on the data and a value based on physical laws, and the learning generates more than twice the predetermined number of data in the generator, and of the generated data, the predetermined number of data with the smallest error and the predetermined number of data with the largest error are used.

(8) In the learning method described in (6) above, the teacher data is data generated by the generator that is determined to be true by the determination means.

(9) In the learning method described in (6) above, when the condition is given as a range of numerical values, the pre-learning is performed using a predetermined number of numerical values included in the range of numerical values as a condition.

(10) In the learning method described in any one of (1) to (9) above, the determination means determines that the data is true if the data is equal to or less than an upper limit of a reference value and equal to or greater than a lower limit of the reference value, and the upper limit is lowered as the learning of the learning model progresses, and the lower limit is raised as the learning of the learning model progresses.

(11) A learning method according to any one of (1) to (9) above, wherein the determination means includes in the determination result the difference between the data and a reference value.

(12) An information processing system consisting of at least one device, the information processing system having at least one processor capable of executing a program to perform each of the steps described in any one of (1) to (11) above.

(13) A program that causes at least one computer to execute each step of the learning method described in any one of (1) to (11) above.

(14) A learning model that is trained by each step of the learning method described in any one of (1) to (11) above.
Of course, this is not the case.

Finally, although various embodiments of the present invention have been described, these are presented as examples and are not intended to limit the scope of the invention. The novel embodiments can be embodied in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. The embodiments and their modifications are within the scope and spirit of the invention, and are included in the scope of the invention and its equivalents as set forth in the claims.

1: Information processing device 11: Processing unit 12: Memory unit 13: Temporary memory unit 14: External device connection unit 15: Communication unit 16: Communication bus 110: Learning device 111: Condition reception unit 112: Learning unit 113: Output unit 114: Reception unit 120: Determination device 121: Determination unit 1121: Generator 1122: Classifier 1123: Comparison unit 1124: Provision unit

Claims (14)

A learning method executed by an information processing system, comprising:
The method includes an output step, a reception step, and a learning step,
In the output step, data generated by the learning model during the learning process, depending on a given condition or without the given condition, is output;
In the receiving step, an input of a determination result of whether the data is true or false by a determination means is received,
In the learning step, learning of the learning model is progressed based on the determination result. 2. The learning method according to claim 1,
The determining means determines the authenticity of the data based on a physical law. 3. The learning method according to claim 2,
The determining means determines whether the condition of the data is true or false. 2. The learning method according to claim 1,
A comparison step is provided,
The learning model includes a generator and a classifier;
The generator generates the data;
The identifier identifies whether the data is genuine or false,
In the comparing step, the judgment result is compared with a classification result by the classifier;
In the learning step, learning of the generator and the discriminator is progressed based on a result of the comparison in the comparison step, and the learning includes:
training the generator to generate data that the classifier classifies as true;
a learning method for causing the classifier to learn to make the same classification as the judgment result of the judgment means as to whether the data is true or false. 5. The learning method according to claim 4,
In the learning step, prior to learning based on a result of the comparison in the comparison step, pre-learning of the generator is performed using teacher data, and the pre-learning includes:
The learning method comprises operating the generator as a neural network using the training data. 5. The learning method according to claim 4,
In the learning step, prior to learning based on a result of the comparison in the comparing step, pre-learning of the generator and the classifier is carried out using teacher data, and the pre-learning includes:
training the generator to generate data that the classifier classifies as true;
A learning method for training the classifier to classify the training data as true. 5. The learning method according to claim 4,
the determining means calculates an error between a value calculated based on the data and a value based on physical laws,
The learning method includes generating more than twice a predetermined number of pieces of data using the generator, and using, from the generated data, the predetermined number of pieces of data with a small error and the predetermined number of pieces of data with a large error. 7. The learning method according to claim 6,
The training data is data generated by the generator and determined to be true by the determination means. 7. The learning method according to claim 6,
a learning method in which, when the condition is given as a range of numerical values, the pre-learning is performed using a predetermined number of numerical values included in the range of numerical values as a condition. 2. The learning method according to claim 1,
the determination means determines the data to be true when the data is equal to or less than an upper limit of a reference value and equal to or greater than a lower limit of the reference value;
The upper limit value is lowered as the learning of the learning model progresses,
The lower limit is increased as the learning of the learning model progresses. 2. The learning method according to claim 1,
The determination means includes a difference between the data and a reference value in the determination result. An information processing system comprising at least one device,
An information processing system comprising at least one processor capable of executing a program to perform each step according to any one of claims 1 to 11. A program,
A program for causing at least one computer to execute each step of the learning method according to any one of claims 1 to 11. A learning model,
A learning model trained by each step of the learning method according to any one of claims 1 to 11.