Model generation device, model generation method, and program

Abstract

To provide a model generation device, a model generation method, and a program capable of generating a highly accurate model.SOLUTION: A model generation device according to the present disclosure includes: a model execution unit that acquires a predicted value of a deep learning model obtained by inputting an input value of a learning dataset used for learning the deep learning model into the deep learning model; a data comparison unit that, on the basis of a ratio of a third term, which is a value based on a difference between a correct value and a predicted value, to a reference term that includes a first term, which is a value based on the correct value in the learning data set, and a second term, which is a value based on a difference between the correct value and the predicted value, calculates a relative error index indicating a difference between a correct value and a predicted value; and a parameter correction unit that, on the basis of the relative error index, corrects a parameter of the deep learning model.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to a model generation device, a model generation method, and a program. [Background technology]

[0002] Various methods have been proposed as training methods for machine learning models such as neural networks. For example, there is a method in which the values ​​output from a neural network are compared with the values ​​of ground truth data, and the parameters of the neural network are modified to minimize a loss function that indicates the difference between the output values ​​and the ground truth data.

[0003] For example, Patent Document 1 below discloses a neural network learning device that adds a regularization term to the loss function of a neural network to make it converge to one true parameter. In this learning device, adding a function consisting of parameters as a regularization term prevents overlearning of input learning data. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 8-202674 Summary of the Invention [Problem to be solved by the invention]

[0005] However, when training as in Patent Document 1, the error in the smallest regions may become relatively excessive or the error may diverge, making it impossible to train properly and potentially making it impossible to build a highly accurate machine learning model.

[0006] In view of the above-described problems, the present disclosure aims to provide a model generation device, a model generation method, and a program that are capable of generating a highly accurate model. [Means for solving the problem]

[0007] In order to solve the above-mentioned problems and achieve the objective, the model generation device of the present disclosure includes: a model execution unit that acquires a predicted value of the deep learning model obtained by inputting input values ​​of a training dataset used for training the deep learning model into the deep learning model; a data comparison unit that calculates a relative error index indicating the difference between the correct value and the predicted value based on the ratio of a third term, which is a value based on the difference between the correct value and the predicted value, to reference terms including a first term, which is a value based on a correct value of the training dataset, and a second term, which is a value based on the difference between the correct value and the predicted value; and a parameter correction unit that corrects parameters of the deep learning model based on the relative error index.

[0008] In order to solve the above-mentioned problems and achieve the objective, the model generation method of the present disclosure includes the steps of: acquiring a predicted value of a deep learning model obtained by inputting input values ​​of a training dataset used for training the deep learning model into the deep learning model; calculating a relative error index indicating the difference between the correct value and the predicted value based on the ratio of a third term, which is a value based on the difference between the correct value and the predicted value, to reference terms including a first term, which is a value based on a correct value of the training dataset, and a second term, which is a value based on the difference between the correct value and the predicted value; and correcting parameters of the deep learning model based on the relative error index.

[0009] In order to solve the above-mentioned problems and achieve the objectives, the program of the present disclosure causes a computer to execute the following steps: acquiring a predicted value of the deep learning model obtained by inputting input values ​​of a training dataset used for training the deep learning model into the deep learning model; calculating a relative error index indicating the difference between the correct value and the predicted value based on the ratio of a third term, which is a value based on the difference between the correct value and the predicted value, to reference terms including a first term, which is a value based on a correct value of the training dataset, and a second term, which is a value based on the difference between the correct value and the predicted value; and correcting parameters of the deep learning model based on the relative error index. [Effects of the Invention]

[0010] According to the present disclosure, it is possible to provide a model generation device, a model generation method, and a program that are capable of generating a highly accurate model. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram illustrating an example configuration of a model generation system according to the present disclosure. [Figure 2] FIG. 2 is a diagram illustrating an example of the configuration of a model generating device according to the present disclosure. [Figure 3] FIG. 3 is a diagram illustrating an example of information stored in a trained model storage unit of the model generation device according to the present disclosure. [Figure 4] FIG. 4 is a diagram illustrating an example of information stored in a learning data storage unit of a model generation device according to the present disclosure. [Figure 5] FIG. 5 is a diagram illustrating an example of the configuration of the data comparison unit of the model generation device according to the present disclosure. [Figure 6] FIG. 6 is a flowchart showing the flow of the model generation method according to the present disclosure. [Figure 7] FIG. 7 is a diagram schematically illustrating a problem used to verify a trained model according to an embodiment. [Figure 8]FIG. 8 is a diagram showing the numerical ranges of the parameters of the trained model according to the present disclosure. [Figure 9] FIG. 9 is a diagram illustrating the layer structure and activation functions of a trained model according to the present disclosure. [Figure 10] FIG. 10 illustrates the mean absolute percentage error of a trained model according to the present disclosure. [Figure 11] FIG. 11 is a diagram illustrating the maximum error of the trained model according to the present disclosure. [Figure 12] FIG. 12 is a diagram illustrating an example of the configuration of an information processing device according to the present disclosure. [Figure 13] FIG. 13 is a diagram illustrating hardware for realizing a model generation device according to the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0012] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. However, the present disclosure is not limited to the embodiments described below.

[0013] (Model Generation System Configuration) First, a model generation system 1 according to the present disclosure will be described with reference to Fig. 1. Fig. 1 is a diagram showing an example configuration of the model generation system according to the present disclosure. As shown in Fig. 1, the model generation system 1 according to the present disclosure includes a model generation device 100, an information processing device 200, and a network N. Below, these components will be briefly described in order.

[0014] The model generation device 100 is an information processing terminal that executes processes such as generating a trained model (deep learning model). The model generation device 100 may be realized by, for example, a computer equipped with the functions of a personal computer (PC), a workstation (WS), or a server. The model generation device 100, for example, accepts input of operation information from a user and performs processing based on the operation information.

[0015] The information processing device 200 is an information processing terminal that executes various types of information processing. The information processing device 200 is used, for example, for executing a trained model. The information processing device 200 may be realized, for example, by a computer having the functions of a PC, a WS, and a server. Note that, as shown in FIG. 1, a plurality of information processing devices 200 may be provided.

[0016] The network N connects the model generation device 100 and the information processing device 200 in a wired or wireless manner so that they can communicate with each other. If the network N is wired, it may be realized by ETHERNET (registered trademark) defined in IEEE802.3. If the network N is wireless, it may be realized by a wireless LAN (Local Area Network) defined in IEEE802.11.

[0017] (Configuration of the model generation device) Next, the configuration of a model generation device 100 according to the present disclosure will be described with reference to Fig. 2. Fig. 2 is a diagram showing an example configuration of a model generation device according to the present disclosure. As shown in Fig. 2, the model generation device 100 according to the present disclosure includes a communication unit 110, a storage unit 120, a control unit 130, a sound output unit 140, an input unit 150, and a display unit 160. These components will be described in order below.

[0018] The communication unit 110 is responsible for transmitting and receiving various information to and from external devices via wired or wireless connections. In the wired case, it may be realized by a network interface card (NIC) equipped with an interface such as a wired LAN terminal. In the wireless case, it may be realized by a wireless LAN defined by IEEE802.11 or the like.

[0019] The storage unit 120 is a storage device that stores various types of information. The storage unit 120 includes a main storage device and an auxiliary storage device. The main storage device may be realized by a semiconductor memory element such as a random access memory (RAM), a read only memory (ROM), or a flash memory. The auxiliary storage device may be realized by a hard disk, a solid state drive (SSD), an optical disk, or the like.

[0020] 3, the storage unit 120 includes a model storage unit 121 and a learning dataset storage unit 122. The information stored in these components will be explained below in order.

[0021] The model storage unit 121 stores information related to the deep learning model. A deep learning model is a model (program) that, when an input value is input, performs a calculation using set parameters and the input value and outputs a predicted value. The parameters used by the deep learning model are learned and set using a machine learning technique. Any deep learning model may be used as the deep learning model. Furthermore, the deep learning model according to this embodiment may be a model that performs any task, but is preferably used for the task of predicting a physical phenomenon. Furthermore, the deep learning model according to this embodiment is preferably used for a task in which the possible values ​​of the predicted value range from negative values ​​to positive values ​​across zero.

[0022] An example of information stored in the model storage unit 121 will now be described with reference to Fig. 3. Fig. 3 is a diagram illustrating an example of information stored in the model storage unit of the model generation device according to the present disclosure.

[0023] As shown in FIG. 3, the model storage unit 121 stores information related to the items "model ID" and "model data".

[0024] A "model ID" is an identifier that identifies a deep learning model and is represented by a string or number. "Model data" is data for the deep learning model identified by the "model ID." The deep learning model may be a deep neural network (DNN). "Model data" includes various information, such as connection information about how nodes included in each of the multiple layers that make up the deep neural network are connected to each other, and connection coefficients (parameters) that are multiplied by numerical values ​​input and output between connected nodes.

[0025] That is, FIG. 3 shows an example in which model data "MDT#1" of a deep learning model identified by a model ID "M#1" is stored.

[0026] Note that the information stored in the model storage unit 121 is not limited to information relating to the items "model ID" and "model data", and may store information relating to any other deep learning model.

[0027] The training dataset storage unit 122 stores information related to a training dataset used for training a deep learning model. The training dataset includes input values ​​to be input to the deep learning model and correct values ​​that are to be trained as desirable output values ​​when an actual task is executed using the input values. The training dataset includes multiple sets of datasets that include input values ​​and correct values ​​corresponding to the input values. An example of information stored in the training dataset storage unit 122 will now be described with reference to FIG. 4. FIG. 4 is a diagram illustrating an example of information stored in the training dataset storage unit of the model generation device according to the present disclosure.

[0028] As shown in FIG. 4, the learning data set storage unit 122 stores information relating to the items "learning data set ID," "input data," and "correct answer data."

[0029] The "learning dataset ID" is an identifier that identifies a learning dataset that includes input data (input values) and correct answer data (correct answer values), and is represented by a string or a number. "Input data" refers to the data in the learning dataset that is input to a deep learning model. "Correct answer data" refers to the data in the learning dataset that is trained as the correct answer by a deep learning model. In this embodiment, since the input values ​​and correct answer values ​​are numerical, the input data and correct answer data are also numerical.

[0030] That is, Figure 4 shows an example in which input data "IPDT#1" and correct answer data "CRDT#1" are combined and stored as a training dataset identified by the training dataset ID "TRID#1".

[0031] The information stored in the training dataset storage unit 122 is not limited to information relating to the items "training dataset ID," "input data," and "correct answer data," and may store information relating to any other training dataset.

[0032] The control unit 130 is a controller that executes various types of arithmetic processing and processing for implementing functions. The control unit 130 is realized by a CPU (Central Processing Unit), an MPU (Micro Processing Unit), or the like executing various programs stored in the storage unit 120 using RAM as a working area. The control unit 130 may also be realized by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

[0033] 3, the control unit 130 includes a data input unit 131, a model execution unit 132, a data comparison unit 133, a parameter correction unit 134, and an error determination unit 135 as functions realized by the execution of a program stored in the storage unit 120, a circuit configuration, etc. The control unit 130 may execute these processes using a single CPU, or may be equipped with multiple CPUs that execute these processes in parallel. The processes executed by these units will be described in order below.

[0034] (Data entry section) The data input unit 131 acquires a training dataset used for training the deep learning model, and inputs input values ​​(input data) from the training dataset to the deep learning model. The data input unit 131 inputs correct values ​​(correct data) from the acquired training dataset to the correct value input unit 1331 of the data comparison unit 133. For example, the data input unit 131 acquires a training dataset stored in the training dataset storage unit 122. As described above, the training dataset storage unit 122 stores a training dataset in which input values ​​and correct values ​​are combined. The data input unit 131 may also acquire a training dataset from an external storage device via the communication unit 210. After acquiring the training dataset stored in the training dataset storage unit 122, the data input unit 131 inputs the input values ​​from the acquired training dataset to the trained model.

[0035] (Model Execution Department) The model execution unit 132 acquires a predicted value of the deep learning model obtained by inputting input values ​​to the deep learning model. The model execution unit 132 acquires a predicted value of the deep learning model for each of multiple input values ​​included in the training dataset. In this embodiment, the model execution unit 132 executes the deep learning model to acquire the predicted value of the deep learning model. Specifically, the model execution unit 132 executes the deep learning model using the input values ​​in the training dataset to acquire the predicted value. That is, the model execution unit 132 inputs the input values ​​to the deep learning model stored in the model storage unit 121, causes the deep learning model to perform calculations based on the input values ​​and set parameters, and acquires a predicted value, which is an output value of the calculation result from the deep learning model. The model execution unit 132 similarly executes the deep learning model for each of the multiple input values ​​included in the training dataset to acquire the predicted value. After acquiring the predicted value of the deep learning model, the model execution unit 132 transmits the acquired predicted value of the deep learning model to the data comparison unit 133, which will be described later. In addition, the model execution unit 132 is not limited to executing the deep learning model itself to obtain the predicted value, but may also obtain the predicted value by receiving the predicted value calculated based on the input value and parameters using the deep learning model from an external device.

[0036] (Data comparison section) The data comparison unit 133 calculates a relative error index indicating the difference between the correct value and the predicted value based on the ratio of the third term, which is a value based on the difference between the correct value and the predicted value, to a base term including a first term, which is a value based on the correct value, and a second term, which is a value based on the difference between the correct value and the predicted value. The relative error index is an index indicating the difference between the correct value and the predicted value, and the parameters of the deep learning model are corrected by the parameter correction unit 134, which will be described later, using this relative error index. The first term, second term, base term, and third term will be described later.

[0037] The configuration of the data comparison unit 133 will now be described with reference to Fig. 5. Fig. 5 is a diagram illustrating an example configuration of the data comparison unit of the model generation device according to the present disclosure. As shown in Fig. 5, the data comparison unit 133 includes a correct value input unit 1331, a predicted value input unit 1332, a correction term calculation unit 1333, and an error index calculation unit 1334. The configuration of these units will be described in order below.

[0038] (Correct answer input section) The correct value input unit 1331 acquires a correct value from the training dataset and inputs the acquired correct value to the correction term calculation unit 1333. That is, the correct value input unit 1331 acquires a correct value from the training dataset that corresponds to an input value input to the deep learning model, and inputs the correct value to the correction term calculation unit 1333, which will be described later. That is, the correct value input to the correction term calculation unit 1333 is the correct value from the training dataset that corresponds to the input value input to the deep learning model.

[0039] (Prediction value input section) The predicted value input unit 1332 acquires the predicted value acquired by the model executing unit 132 (the predicted value of the deep learning model obtained by inputting the input value to the deep learning model), and inputs the acquired predicted value to the correction term calculation unit 1333. That is, the predicted value input to the correction term calculation unit 1333 is the predicted value obtained by inputting the input value of the learning dataset to the deep learning model, and the input value of the learning dataset is associated with the correct answer value.

[0040] (Correction term calculation section) The correction term calculation unit 1333 calculates the second term (correction term). The second term is a value based on the difference between the correct value and the predicted value, and is a dynamic term whose value changes depending on the difference between the correct value and the predicted value. In this embodiment, the correction term calculation unit 1333 calculates the second term by using a value obtained by converting the difference between the correct value and the predicted value into a positive value as a value based on the difference between the correct value and the predicted value. For example, the correction term calculation unit 1333 may use the square root of the difference between the correct value and the predicted value as the value based on the difference between the correct value and the predicted value, or may use the absolute value of the difference between the correct value and the predicted value as the value based on the difference between the correct value and the predicted value. Furthermore, in this embodiment, since multiple sets are set for each of the input value, correct value, and predicted value, the correction term calculation unit 1333 calculates the second term using a value based on the difference between the correct value and the predicted value for each of the associated sets of correct value and predicted value. In this embodiment, the correction term calculation unit 1333 calculates the average value of the values ​​based on the difference between the correct value and the predicted value of each set as the second term.

[0041] The second term is a value based on the difference between the correct value and the predicted value, and may be any term whose value changes depending on the difference between the correct value and the predicted value. In this embodiment, the correction term calculation unit 1333 calculates the root mean squared error (RMSE) for each of the correct value and predicted value of each associated set as the second term. Note that the RMSE is expressed by the following formula (1).

[0042]

number

[0043] Here, in equation (1), y true is the correct value of the training dataset, y pred is the predicted value of the deep learning model when the input values ​​from the training dataset are input to the deep learning model, and n is the total number of input values ​​input to the deep learning model.

[0044] That is, the correction term calculation unit 1333 calculates the RMSE as the second term from the predicted value of the deep learning model and the correct value of the learning dataset based on the above formula (1). As shown in formula (1), the correction term calculation unit 1333 calculates the difference between the correct value and the predicted value of each associated set, calculates the square value, divides the sum by the number of datasets, and takes the square root to calculate the second term.

[0045] However, the second term is not limited to RMSE. For example, the correction term calculation unit 1333 may calculate the mean absolute error (MAE) for each of the correct value and the predicted value of each associated set as the second term. MAE is expressed by the following formula (2).

[0046]

number

[0047] That is, in this example, the correction term calculation unit 1333 calculates the RMSE as the second term from the predicted value of the deep learning model and the correct value of the learning dataset based on the above formula (2). Note that, as shown in formula (2), the correction term calculation unit 1333 calculates the absolute value of the difference between the correct value and the predicted value of each associated set, and divides the sum by the number of datasets to calculate the second term.

[0048] (Error index calculation part) The error index calculation unit 1334 calculates a relative error index indicating the difference between the correct value and the predicted value based on the ratio of the third term to a base term including the first and second terms.

[0049] As described above, the first term is a value based on the correct value, and can be said to be a value based on the correct value but not on the predicted value. Therefore, the value of the first term changes depending on the correct value, but does not change depending on the predicted value. The error index calculation unit 1334 may use the correct value itself as the first term, but in this embodiment, a value obtained by converting the correct value into a positive value is used as the first term. For example, the correction term calculation unit 1333 may calculate the square of the correct value as the first term, or may calculate the absolute value of the correct value as the first term. Note that while the second term is a value based on the difference between the correct value and the predicted value of each set, the first term is a value based on the correct value of one set.

[0050] The second term is a value calculated by the above-mentioned correction term calculation unit 1333. The reference term is a term including the first and second terms. The reference term may be any term including the first and second terms, but in this embodiment, the sum of the first and second terms is used as the reference term.

[0051] As described above, the third term is a value based on the difference between the correct value and the predicted value, and is a dynamic term whose value changes depending on the difference between the correct value and the predicted value. In this embodiment, the correction term calculation unit 1333 calculates the second term by using a value obtained by converting the difference between the correct value and the predicted value into a positive value as a value based on the difference between the correct value and the predicted value. The correction term calculation unit 1333 may use the difference between the correct value and the predicted value itself as the third term, but in this embodiment, the correction term calculation unit 1333 uses the value obtained by converting the difference between the correct value and the predicted value into a positive value as the third term. For example, the correction term calculation unit 1333 may calculate the square of the difference between the correct value and the predicted value as the third term, or may calculate the absolute value of the difference between the correct value and the predicted value as the third term. Note that while the second term is a value based on the difference between the correct value and the predicted value of each set, the third term is a value based on the difference between the correct value and the predicted value of one set.

[0052] In this embodiment, the error index calculation unit 1334 calculates the relative error index by using the reference term, which is the sum of the first term and the second term, as the denominator and the third term as the numerator. Specifically, the error index calculation unit 1334 calculates the reference term and the third term for each of the associated correct value and predicted value of each set, and calculates the relative error index based on the ratio of the third term to the reference term of each set. In this embodiment, the error index calculation unit 1334 calculates the average value of the ratio of the third term to the reference term of each set as the relative error index. Note that in this case, since the second term is a value based on the difference between the correct value and the predicted value of each set, a common value is used as the second term when calculating the ratio of each set.

[0053] The relative error index may be any value based on the ratio of the third term to the reference term, but in this embodiment, the error index calculation unit 1334 calculates the square root of the mean squared percentage error with the second term introduced into the denominator as the relative error index. Specifically, the error index calculation unit 1334 calculates the root mean squared percentage error (RMSPE) with the second term, which is the RMSE, introduced into the denominator as the relative error index. Note that the root mean squared percentage error (RMSPE) is the value of the square root of the mean squared percentage error. The root mean squared percentage error (RMSPE) with the second term introduced into the denominator is ** ) is expressed by the following equation (3).

[0054]

number

[0055] That is, based on the above formula (3), the error index calculation unit 1334 calculates the relative error index by dividing the sum of the ratio of the third term (in formula (3), the square of the difference between the correct value and the predicted value) to the sum of the first term (in formula (3), the square of the correct value) and the second term (in formula (3), the RMSE) for each set by the number of data sets.

[0056] However, the relative error index is RMSPE **For example, the error index calculation unit 1334 may calculate the mean absolute percentage error (MAPE) with the second term introduced into the denominator as the relative error index. Specifically, the error index calculation unit 1334 calculates the mean absolute percentage error (MAPE) with the second term, which is the MAE, introduced into the denominator as the relative error index. ** ) is expressed by the following equation (4).

[0057]

number

[0058] That is, in this example, the error index calculation unit 1334 calculates the relative error index based on the above formula (4) by dividing the sum of the ratio of the third term (the absolute value of the difference between the correct value and the predicted value in formula (4)) to the sum of the first term (the absolute value of the correct value in formula (4)) and the second term (the MAE in formula (4)) for each set by the number of data sets.

[0059] (Parameter modification section) The parameter correction unit 134 corrects the parameters of the deep learning model based on the relative error index. Specifically, the parameter correction unit 134 searches for parameters that minimize a loss function represented by the relative error index calculated by the error index calculation unit 1334, and corrects the parameters of the deep learning model using the searched parameters. Note that the parameter correction unit 134 may use, as an optimization algorithm for searching for parameters, gradient descent, stochastic gradient descent (SDG), mini-batch learning SDG, momentum, adaptive gradient algorithm (AdaGrad), RMSProp, Adam, or the like.

[0060] (Error judgment section) The error determination unit 135 determines whether the relative error index of the deep learning model after the parameters have been corrected is the minimum value. If the error determination unit 135 determines that the relative error index of the deep learning model after the parameters have been corrected is the minimum value, the parameter correction is completed, and a deep learning model incorporating the corrected parameters is generated.

[0061] If the error determination unit 135 determines that the relative error index of the deep learning model after correcting the parameters is not the minimum value, the process returns to the processing of the parameter correction unit 134, where the parameters are adjusted again, and the relative error index is calculated again using the predicted value obtained by inputting the input data from the learning dataset into the deep learning model by the data input unit 131, thereby correcting the parameters. That is, the error determination unit 135 repeats the parameter correction until it determines that the relative error index of the deep learning model is the minimum value.

[0062] (Sound output section) The sound output unit 140 outputs various sounds. For example, when processing in the model generation device 100 is completed, the sound output unit 140 may output a sound notifying that the processing is completed. The sound output unit 140 may be a speaker, which converts an electrical signal into sound using a diaphragm. That is, the speaker vibrates the diaphragm with a predetermined amplitude and frequency based on a control command given by the electrical signal, thereby vibrating the air in contact with the diaphragm and outputting sound.

[0063] (Input section) Various types of operation information are input from the user to the input unit 150. For example, the input unit 150 may accept various operations from the user via a display surface (for example, the display unit 160) using a touch panel. Alternatively, the input unit 150 may accept various operations from the user using various buttons, a keyboard, or a mouse.

[0064] (Display) The display unit 160 displays various types of information. For example, the display unit 160 may display a GUI (Graphical User Interface) for receiving operations related to various processes from the user, or the results of various processes. The display unit 160 may be realized by a liquid crystal display, an organic EL (Electro Luminescence) display, a micro LED (Light Emitting Diode) display, or the like. The display unit 160 may also be a touch panel of various types, such as a capacitive type.

[0065] (About model generation method) Next, a model generation method according to the present disclosure will be described with reference to Fig. 6. Fig. 6 is a flowchart showing the flow of the model generation method according to the present disclosure. Below, the model generation method according to the present disclosure will be described along the flow shown in Fig. 6.

[0066] First, the model generation device 100 sets parameters of the deep learning model (step S101). Next, the model generation device 100 calculates a predicted value using the deep learning model (step S102). Next, the model generation device 100 calculates a second term (correction term) based on the predicted value and the correct value (step S103). Next, the model generation device 100 calculates a relative error index (step S104). Next, the model generation device 100 adjusts parameters of the deep learning model (step S105). Next, the model generation device 100 determines whether the relative error index is a minimum value (step S106). Next, the model generation device 100 determines whether the minimum value has not been updated an arbitrary number of times (step S107). If the minimum value has not been updated an arbitrary number of times in step S107 (step S107: Yes), the model generation device 100 ends the processing of the model generation method. Note that the arbitrary number of times may be, for example, 100 times.

[0067] In step S107, if the minimum value is updated before reaching the predetermined number of times (step S107: No), the model generation device 100 returns to step S102 and executes the subsequent processes. That is, after it is determined that the relative error index is the minimum value, the deep learning model parameters are adjusted repeatedly until the minimum value is not updated an predetermined number of times.

[0068] (effect) Here, when setting parameters for a deep learning model through machine learning, the error between a known correct value and a predicted value obtained by inputting an input value into the deep learning model may be used. However, when machine learning is performed using the error between the correct value and the predicted value in this manner, the error may diverge, potentially preventing proper machine learning. For example, if the error is calculated using the correct value as the denominator and the difference between the correct value and the predicted value as the numerator, the error may diverge if the correct value is zero. Furthermore, the error in the minimum region (i.e., when the correct value is small) may be excessively large relative to the error in other regions, resulting in an unbalanced weighting of the error and preventing proper machine learning. In contrast, in this embodiment, the error is calculated based on the ratio of the third term based on the difference between the correct value and the predicted value to a base term including the first term based on the correct value and the second term based on the difference between the correct value and the predicted value, and the parameters are corrected accordingly. In this way, by calculating the error using the second term as the denominator, the possibility of the denominator becoming zero is reduced, thereby suppressing error divergence. Furthermore, by introducing the second term as the denominator to calculate the error, the value of the denominator in the minimal region is increased, preventing the error in the minimal region from becoming relatively excessive, thereby suppressing bias in the weighting of the error. Therefore, according to this embodiment, a deep learning model can be appropriately machine-learned to generate a deep learning model with high accuracy (small error).

[0069] Furthermore, in this embodiment, the second term is calculated based on a value based on the difference between the correct value and the predicted value for each set of associated correct values ​​and predicted values. By using a value based on the difference between each set as the second term in this way, the possibility that the denominator will be 0 can be more appropriately reduced and the value of the denominator in the minimum region can be more appropriately increased. Therefore, a deep learning model can be more appropriately machine-learned to generate a deep learning model with high accuracy (small error).

[0070] (Example) Next, an example of verifying the accuracy of a trained model generated by the model generation device 100 according to the present disclosure will be described.

[0071] In the example, the problem was to predict the displacement of the nodes of a one-dimensional cantilever beam shown in FIG. 7 when a load is applied to the end point of the cantilever beam. FIG. 7 is a diagram schematically illustrating the problem used to verify the trained model in the example. FIG. 7 shows a cantilever beam model having nodes N2 and N3 supported at node N1, which is a fixed point, element E1 connecting node N1 and node N2, and element E2 connecting node N2 and node N3. The cantilever beam model shows that a load F is applied in the tensile direction at node N3. Element E1 has a cross-sectional area set to A1, a length set to L1, and an elastic modulus set to E1. Element E2 has a cross-sectional area set to A2, a length set to L2, and an elastic modulus set to E2.

[0072] The training dataset for training the deep learning model was created by performing FEM (Finite Element Method) analysis using multiple combinations of parameters shown in FIG. 8. FIG. 8 is a diagram showing the numerical ranges of the training dataset used in training for verifying the deep learning model according to the present disclosure. FIG. 8 shows the numerical ranges of the cross-sectional areas A1 and A2, lengths L1 and L2, and elastic moduli E1 and E2, with their minimum and maximum values. That is, within these numerical ranges, each variable was assigned, and FEM analysis was performed. The obtained analysis results were used as the correct values ​​for the training dataset, and the data input into the FEM analysis were used as the input values ​​for the training dataset.

[0073] In other words, a deep learning model trained using such a training dataset can be said to be a surrogate model. A surrogate model is a method of replacing physical simulation with machine learning. A surrogate model is a deep learning model trained to receive the same information as a CAE (Computer Aided Engineering) model as input and produce the same output as the CAE model. For example, if the surrogate model is trained to learn the relationship between the input and output of an FEM model, it will output a load-displacement curve, strain distribution, etc.

[0074] Next, the layer structure and activation function of the deep learning model used in the verification will be described with reference to FIG. 9. FIG. 9 is a diagram showing the layer structure and activation function of a trained model according to an embodiment. As shown in FIG. 9, the deep learning model used in the verification includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer. The input layer has 8 nodes, the first hidden layer has 64 nodes, the second hidden layer has 256 nodes, the third hidden layer has 64 nodes, and the output layer has 2 nodes. The activation function used in the first hidden layer, the second hidden layer, and the third hidden layer is a rectified linear unit (ReLu) function, and the activation function used in the output layer is an identity function. The activation function is a function that determines the final output of a single neuron in a neural network using the sum of multiple nodes as input.

[0075] Using the learning datasets described above, machine learning was performed using different methods to prepare a deep learning model of Example 1, a deep learning model of Example 2, and a deep learning model of Comparative Example 1. In Example 1, the RMSPE of Equation (3) in the above-described embodiment was ** In Example 2, the MAPE of Equation (4) in the above embodiment was used as the relative error index. ** In Comparative Example 1, a deep learning model was trained by machine learning using the RMSE of Equation (1) in the above-described embodiment as a relative error index. In another comparative example, a deep learning model was trained by machine learning using the RMSE of Equation (3) ** (the formula obtained by removing the RMSE from the denominator), MAPE (the MAPE in equation (4) ** The deep learning model was trained using the relative error index (the formula obtained by removing MAE from the denominator of the above), but the denominator of the relative error index diverged, making it impossible to perform the machine learning.

[0076] Figure 10 shows the mean absolute percentage error of the deep learning model used in the verification. As shown in Figure 10, using the machine-learned deep learning models of Examples 1 and 2 and Comparative Example 1, predicted values ​​for each input value were calculated, and the root mean square error (RMSE) between the correct value and the predicted value was calculated. As shown in Figure 10, the MAPE of Comparative Example 1 was 28, and the RMSE of Example 1 was 10. ** The MAPE for Example 1 was 6.3, and the MAPE for Example 2 was 5.0. That is, in the case of a trained model generated by correcting the parameters of a deep learning model using the terms in Equation (2) or Equation (4), i.e., the term with the second term as the denominator, as a relative error index, as in Examples 1 and 2, it was confirmed that the MAPE can be reduced and high accuracy can be achieved compared to the case where the term with the second term as the denominator is not used as a relative error index, as in the comparative example.

[0077] FIG. 11 is a diagram showing the mean absolute error of the deep learning model used in the verification. Furthermore, as shown in FIG. 11, using the machine-learned deep learning models of Examples 1 and 2 and Comparative Example 1, predicted values ​​for each input value were calculated, and the maximum error (ME; Max Error) between the correct value and the predicted value was calculated. As shown in FIG. 11, the maximum error (ME) of Comparative Example 1 was 0.29, the maximum error (ME) of Example 1 was 0.26, and the maximum error (ME) of Example 2 was 1.03. That is, as in Example 1, RSMPE ** When using RSMPE as in Example 1, the maximum error (ME) can be made smaller than in Example 2 and Comparative Example 1. ** It is found that it is more preferable to use

[0078] (Configuration of information processing device) Next, the configuration of the information processing device 200 according to the present disclosure will be described with reference to Fig. 12. Fig. 12 is a diagram showing an example configuration of the information processing device according to the present disclosure. As shown in Fig. 12, the information processing device 200 according to the present disclosure includes a communication unit 210, a storage unit 220, a control unit 230, a sound output unit 240, an input unit 250, and a display unit 260.

[0079] The data input unit 231 acquires data to be input to the trained model and inputs the acquired data to the trained model. For example, the data input unit 231 acquires data input by a user via the input unit 250. The data input unit 231 may also acquire data from an external device via the communication unit 210. After acquiring the data to be input to the trained model, the data input unit 231 inputs the acquired data to the trained model.

[0080] The model execution unit 232 executes the trained model based on the input data. That is, the model execution unit 232 reads the trained model from the model storage unit 221, inputs the data, and outputs a predicted value from the trained model.

[0081] The result output unit 233 displays the predicted value, which is the output of the trained model. That is, the result output unit 233 displays the predicted value, which is the output of the trained model, on the display unit 260.

[0082] According to the information processing device 200 described above, it is possible to execute a trained model with reduced error that has been generated by correcting the parameters of the deep learning model using the relative error index.

[0083] (Hardware configuration) The model generation device 100 according to the above-described embodiment is realized by, for example, a computer 1000 configured as shown in Fig. 13. Fig. 13 is a hardware configuration diagram showing an example of a computer that realizes the functions of the model generation device according to the present disclosure. The computer 1000 is connected to an output device 1010 and an input device 1020, and has a configuration in which a calculation device 1030, a primary storage device 1040, a secondary storage device 1050, an output IF (Interface) 1060, an input IF 1070, and a network IF 1080 are connected via a bus 1090.

[0084] The arithmetic device 1030 operates based on programs stored in the primary storage device 1040 and secondary storage device 1050, programs read from the input device 1020, and the like, and executes various processes. The primary storage device 1040 is a memory device, such as a RAM, that temporarily stores data used by the arithmetic device 1030 for various calculations. The secondary storage device 1050 is a storage device that stores data used by the arithmetic device 1030 for various calculations and various databases, and is realized by a ROM (Read Only Memory), an HDD (Hard Disk Drive), a flash memory, or the like.

[0085] The output IF 1060 is an interface for transmitting information to be output to an output device 1010 that outputs various types of information, such as a monitor or a printer, and is realized by a connector conforming to a standard such as USB (Universal Serial Bus), DVI (Digital Visual Interface), or HDMI (High Definition Multimedia Interface), etc. The input IF 1070 is an interface for receiving information from various input devices 1020, such as a mouse, keyboard, scanner, etc., and is realized by a USB, etc.

[0086] The input device 1020 may be a device that reads information from, for example, an optical recording medium such as a CD (Compact Disc), a DVD (Digital Versatile Disc), or a PD (Phase Change Rewritable Disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), a tape medium, a magnetic recording medium, or a semiconductor memory. The input device 1020 may also be an external storage medium such as a USB memory.

[0087] The network IF 1080 receives data from other devices via the network N and sends it to the arithmetic device 1030, and also transmits data generated by the arithmetic device 1030 to other devices via the network N.

[0088] The arithmetic unit 1030 controls the output device 1010 and the input device 1020 via the output IF 1060 and the input IF 1070. For example, the arithmetic unit 1030 loads a program from the input device 1020 or the secondary storage device 1050 onto the primary storage device 1040 and executes the loaded program.

[0089] For example, when the computer 1000 functions as the model generating device 100, the arithmetic unit 1030 of the computer 1000 realizes the functions of the control unit 130 of the model generating device 100 by executing a program loaded onto the primary storage device 1040.

[0090] (Composition and Effects) The model generation device 100 according to the first aspect includes a model execution unit 132 that acquires a predicted value of the deep learning model obtained by inputting input values ​​of a training dataset used for training the deep learning model into the deep learning model; a data comparison unit 133 that calculates a relative error index indicating the difference between the correct value and the predicted value based on the ratio of a third term, which is a value based on the difference between the correct value and the predicted value, to reference terms including a first term, which is a value based on the correct value of the training dataset, and a second term, which is a value based on the difference between the correct value and the predicted value; and a parameter correction unit 134 that corrects parameters of the deep learning model based on the relative error index.

[0091] According to this configuration, by adopting a relative error index as an index for correcting the parameters of a deep learning model, learning can be performed to reduce the relative error even for data with small physical quantities, thereby providing a model generation device 100 that can generate highly accurate models.

[0092] The model generating device 100 according to the second aspect is the model generating device 100 according to the first aspect, and the data comparing unit 133 of the model generating device 100 includes: a correction term calculating unit 1333 that calculates the second term based on a value that is based on the difference between the correct value and the predicted value for each set of associated correct value and predicted value; and an error index calculating unit 1334 that calculates a relative error index using a reference term obtained by adding the second term to the first term as the denominator and a third term as the numerator.

[0093] According to this configuration, by introducing a correction term into the denominator of the relative error index, a predetermined value is added to the value of the denominator even when the correct value is close to 0, thereby preventing divergence. Therefore, it is possible to provide a model generation device 100 that can generate highly accurate models.

[0094] The model generating device 100 according to the third aspect is the model generating device 100 according to the first or second aspect, in which the correction term calculation unit 1333 of the model generating device 100 calculates, as the second term, the root mean square error between the predicted value and the correct value, and the error index calculation unit 1334 calculates, as the relative error index, the square root of the mean square percentage error with the second term introduced into the denominator.

[0095] This configuration allows the square root of the mean square percentage error, in which the root mean square error between the predicted value and the correct value is introduced into the denominator, to be calculated as the relative error index, thereby providing a model generation device 100 that can generate highly accurate models.

[0096] The model generation device 100 according to the fourth aspect is the model generation device 100 according to any one of the first to third aspects, and the parameter correction unit 134 of the model generation device 100 searches for parameters that minimize a loss function represented by a relative error index, and corrects the parameters of the deep learning model using the searched parameters.

[0097] This configuration allows a search for parameters that minimize the relative error index with a correction term in the denominator, and the searched parameters are used to correct the parameters of the deep learning model, thereby providing a model generation device 100 that can generate highly accurate models.

[0098] The model generation device 100 according to the fifth aspect is the model generation device 100 according to any one of the first to fourth aspects, and further includes an error determination unit 135 that determines whether the relative error index of the deep learning model after parameter correction is the minimum value.

[0099] This configuration makes it possible to determine whether the relative error index of a deep learning model using parameters corrected by previous learning is at a minimum value. This allows the parameters to be corrected appropriately. Therefore, it is possible to provide a model generation device 100 that can generate a model with reduced error by appropriately adjusting the parameters.

[0100] The model generation method according to the first aspect includes the steps of: acquiring a predicted value of the deep learning model obtained by inputting input values ​​from a training dataset used for training the deep learning model into the deep learning model; calculating a relative error index indicating the difference between the correct value and the predicted value based on the ratio of a third term, which is a value based on the difference between the correct value and the predicted value, to a reference term including a first term, which is a value based on the correct value from the training dataset, and a second term, which is a value based on the difference between the correct value and the predicted value; and correcting parameters of the deep learning model based on the relative error index.

[0101] According to this configuration, by adopting a relative error index as an index for correcting the parameters of a deep learning model, learning can be performed to reduce the relative error even for data with small physical quantities. Furthermore, by introducing a correction term into the denominator of the relative error index, a predetermined value is added to the denominator value even when the correct value is close to 0, thereby preventing divergence. Therefore, a model generation method capable of generating highly accurate models can be provided.

[0102] The program according to the first aspect causes a computer to execute the following steps: acquiring a predicted value of the deep learning model obtained by inputting input values ​​from a learning dataset used for training the deep learning model into the deep learning model; calculating a relative error index indicating the difference between the correct value and the predicted value based on the ratio of the third term, which is a value based on the difference between the correct value and the predicted value, to reference terms including a first term, which is a value based on the correct value from the learning dataset, and a second term, which is a value based on the difference between the correct value and the predicted value; and correcting parameters of the deep learning model based on the relative error index.

[0103] According to this configuration, by adopting a relative error index as an index for correcting the parameters of a deep learning model, learning can be performed to reduce the relative error even for data with small physical quantities. Furthermore, by introducing a correction term into the denominator of the relative error index, a predetermined value is added to the denominator value even when the correct value is close to 0, preventing divergence. Therefore, a program capable of generating a highly accurate model can be provided.

[0104] Although the embodiments of the present invention have been described above, the embodiments are not limited to the contents of these embodiments. Furthermore, the above-described components include those that can be easily imagined by a person skilled in the art, those that are substantially the same, and those that are within the scope of what is called equivalents. Furthermore, the above-described components can be combined as appropriate. Furthermore, various omissions, substitutions, or modifications of the components can be made without departing from the spirit of the above-described embodiments. [Explanation of symbols]

[0105] 100 Model generation device 110 Communications Department 120 Storage section 121 Model Memory Unit 122 Learning dataset storage unit 130 Control Unit 131 Data Entry Section 132 Model Execution Department 133 Data comparison section 1331 Correct answer input section 1332 Prediction value input section 1333 Correction term calculation section 1334 Error index calculation section 134 Parameter Modification Section 135 Error judgment section 140 Sound output unit 150 Input section 160 Display section 200 Information processing device 210 Communications Department 220 Storage section 221 Model Memory Unit 230 Control Unit 231 Data Entry Section 232 Model Execution Department 233 Result output section 240 Sound output section 250 Input section 260 Display N Network

Claims

1. a model execution unit that acquires a predicted value of the deep learning model obtained by inputting input values ​​of a learning dataset used for learning the deep learning model into the deep learning model; a data comparator that calculates a relative error index indicating a difference between the correct value and the predicted value based on a ratio of a third term, which is a value based on a difference between the correct value and the predicted value, to reference terms including a first term, which is a value based on a correct value in the training dataset, and a second term, which is a value based on a difference between the correct value and the predicted value; and A parameter correction unit that corrects parameters of the deep learning model based on the relative error index. Model generation device.

2. The data comparison unit a correction term calculation unit that calculates the second term based on a value based on a difference between a correct value and a predicted value for each set of the associated correct value and the predicted value; an error index calculation unit that calculates the relative error index using the reference term obtained by adding the second term to the first term as a denominator and the third term as a numerator, The model generating device according to claim 1 .

3. the correction term calculation unit calculates, as the second term, a root mean square error between the predicted value and the correct value; the error index calculation unit calculates, as the relative error index, the square root of the mean square percentage error with the second term introduced into the denominator; The model generating device according to claim 2 .

4. The parameter correction unit searches for parameters that minimize a loss function represented by the relative error index, and corrects parameters of the deep learning model using the searched parameters. The model generating device according to any one of claims 1 to 3.

5. An error determination unit that determines whether the relative error index of the deep learning model after parameter correction is a minimum value. The model generating device according to claim 4 .

6. A step of acquiring a predicted value of the deep learning model obtained by inputting input values ​​of a learning dataset used for learning the deep learning model into the deep learning model; calculating a relative error index indicating a difference between the correct value and the predicted value based on a ratio of a third term, which is a value based on a difference between the correct value and the predicted value, to reference terms including a first term, which is a value based on a correct value in the training dataset, and a second term, which is a value based on a difference between the correct value and the predicted value; and modifying parameters of the deep learning model based on the relative error metric. Model generation method.

7. A step of acquiring a predicted value of the deep learning model obtained by inputting input values ​​of a learning dataset used for learning the deep learning model into the deep learning model; calculating a relative error index indicating a difference between the correct value and the predicted value based on a ratio of a third term, which is a value based on a difference between the correct value and the predicted value, to reference terms including a first term, which is a value based on a correct value in the training dataset, and a second term, which is a value based on a difference between the correct value and the predicted value; modifying parameters of the deep learning model based on the relative error metric; to the computer, program.

Images