Information processing method and information processing device

Abstract

To provide an information processing method and an information processing apparatus capable of preventing the accuracy of calculating prediction data from becoming lower than required accuracy.SOLUTION: An information processing apparatus 100 is configured to: calculate a first range A1 which is a data range including multiple pieces of learning input data; calculate a second range A2 narrower than the first range A1 and including input data in the case in which an error between a true value and prediction data calculated by a prediction model 7 is equal to or smaller than a predetermined value; calculate an intermediate range Am included in the first range A1, but not included in the second range A2; output, when the input data is included in the intermediate range Am, an inference result by inputting the input data to a classifier 4; specify a class of the highest likelihood as a highest likelihood class; and output, when first prediction data P1 calculated by the prediction model 7 is not included in the highest likelihood class, second prediction data P2 calculated without using the prediction model 7, as prediction data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an information processing method and an information processing device. [Background technology]

[0002] The information processing device described in Patent Document 1 selects a prediction model for predicting aerodynamic performance based on input data indicating the shape of the vehicle, and outputs the aerodynamic performance value calculated using the selected prediction model as prediction data. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2014-006812 Summary of the Invention [Problem to be solved by the invention]

[0004] However, when the prediction model selected by the information processing device described in Patent Document 1 is used, the accuracy of calculation of prediction data may be lower than the required accuracy depending on the range of input data.

[0005] An object of the present invention is to provide an information processing method and an information processing device that can prevent the accuracy of calculation of predicted data from being lower than the required accuracy. [Means for solving the problem]

[0006] The present invention solves the above problem by calculating a first range including multiple learning input data, calculating a second range that is narrower than the first range and includes input data for which the error between the predicted data calculated by the predictive model and the true value is less than a predetermined value, calculating an intermediate range that is included in the first range but not included in the second range, determining whether the first predicted data calculated by the predictive model is included in the highest likelihood class of the inference results output by the classifier if the first predicted data is not included in the highest likelihood class, and outputting second predicted data calculated without using the predictive model as predicted data. [Effects of the Invention]

[0007] The present invention has the effect of preventing the accuracy of calculation of predicted data from being lower than the required accuracy by outputting second predicted data calculated without using a predictive model as predicted data when the input data is within the intermediate range and the first predicted data calculated by the predictive model does not match the inference result of the classifier. [Brief explanation of the drawings]

[0008] [Figure 1] 1 is a block diagram showing a configuration of an information processing device according to an embodiment of the present invention; [Figure 2] 2 is a diagram schematically showing an example of a first range, a second range, and an intermediate range of input data set by the information processing device shown in FIG. 1 on a two-axis graph. FIG. [Figure 3] 2 is a diagram showing an example of an inference result output by a classifier of the information processing device shown in FIG. 1. [Figure 4] 2 is a flowchart showing the procedure of an information processing method executed by the information processing device shown in FIG. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The information processing device 100 has a prediction model 7, which is a trained model obtained by machine learning using a training dataset D including a plurality of pieces of training input data and training measurement data corresponding to each of the plurality of pieces of training input data. The information processing device 100 outputs prediction data based on the input data. The training measurement data is data measured, detected, or calculated using simulations or experiments in accordance with the training input data. The training measurement data and prediction data are data expressed as scalar values. In the following description of this embodiment, the training input data and input data are vehicle shape data (three-dimensional data or side-view images of the vehicle body), and the training measurement data and prediction data are air resistance coefficients. In addition, when the training input data and input data are vehicle shape data, the training measurement data and prediction data may be lift coefficients. In other words, when the training input data and input data are vehicle shape data, the training measurement data and prediction data are values ​​indicating the aerodynamic performance of the vehicle.

[0010] As shown in FIG. 1, the information processing device 100 includes an input unit 1, a data range setting unit 2, an input data range determination unit 3, a classifier 4, an inference result analysis unit 5, a measurement unit 6, a prediction model 7, a learning unit 8, and an output unit 9. The information processing device 100 also stores a learning dataset D. The information processing device is a computer including a CPU, a ROM, and a RAM. The input unit 1, the data range setting unit 2, the input data range determination unit 3, the classifier 4, the inference result analysis unit 5, the measurement unit 6, the prediction model 7, the learning unit 8, and the output unit 9 are programs that execute the functions of the information processing device 100. The input unit 1, the data range setting unit 2, the input data range determination unit 3, the classifier 4, the inference result analysis unit 5, the measurement unit 6, the prediction model 7, the learning unit 8, and the output unit 9 may be programs on a single piece of hardware, or each may be a program on separate hardware.

[0011] Input data is input to the input unit 1 via an input device (not shown). The input device is, for example, a user interface that can be operated by a user, an imaging device, a detection device, etc. If the input data is three-dimensional data of the shape of a vehicle, the input data is, for example, SDF data or voxel data.

[0012] The data range setting unit 2 includes a first range calculation unit 2a, a second range calculation unit 2b, and an intermediate range calculation unit 2c. The data range setting unit 2 sets a data range for determining whether input data can be handled by the prediction model 7. The first range calculation unit 2a calculates a first range A1, which is a data range including multiple pieces of learning input data in the learning dataset D. The first range A1 is a data range that serves as a criterion for determining whether data input to the input unit 1 cannot be handled by the prediction model 7. The second range calculation unit 2b calculates a second range A2, which is a data range that is narrower than the first range A1 and includes input data for which the error between the predicted data calculated by the prediction model 7 and the true value is equal to or smaller than a predetermined value. Note that the true value in this case is measurement data measured by the measurement unit 6 using a CFD simulation or the like. The second range A2 is a data range that serves as a criterion for determining whether data input to the input unit 1 can be handled by the prediction model 7. Furthermore, the intermediate range calculation unit 2c calculates an intermediate range Am that is included in the first range A1 and not included in the second range A2.

[0013] FIG. 2 schematically illustrates a first range A1, a second range A2, and an intermediate range Am when it is assumed that the distribution of input data can be represented on a graph with two axes (x-axis and y-axis). In the example shown in FIG. 2, the first range A1 is a circular region with a radius of R_all_max that encompasses all of the multiple learning input data in the learning dataset D. The second range A2 is a circular region with a radius of R_sub_max that encompasses input data when the error between the predicted data calculated by the prediction model 7 and the true value is equal to or less than a predetermined value. The intermediate range Am is a region inside the first range A1 and outside the second range A2. When the distribution of input data can be represented on a graph with three axes (x-axis, y-axis, and z-axis), the first range A1 and the second range A2 are spherical with radii R_all_max and R_sub_max, respectively. Furthermore, when the distribution of input data can be represented on a graph with four or more axes, the first range A1 and the second range A2 are hyperspherical with radii R_all_max and R_sub_max, respectively. The first range A1 is not limited to a circular, spherical, or hyperspherical range, as long as it includes all of the learning input data in the learning dataset. The second range A2 is not limited to a circular, spherical, or hyperspherical range, as long as it is narrower than the first range A1 and includes input data for which the error between the predicted data calculated by the prediction model 7 and the true value is equal to or less than a predetermined value. The range outside the first range A1 is a data range that the prediction model 7 cannot handle, and the range inside the second range A2 is a data range that the prediction model 7 can handle.

[0014] The input data range determination unit 3 determines whether or not the input data input to the input unit 1 of the information processing device 100 is included in the intermediate range Am. The input data range determination unit 3 also determines whether or not the input data is included in the first range A1, and if the input data is not included in the first range A1 (if the input data is distributed outside the first range A1), determines that the prediction model 7 cannot handle the input data. The input data range determination unit 3 also determines whether or not the input data is included in the second range A2, and if the input data is included in the second range A2 (if the input data is distributed inside the second range A2), determines that the prediction model 7 can handle the input data.

[0015] On the other hand, if the input data range determination unit 3 determines that "the input data falls within the intermediate range Am," the input data is input to the classifier 4 shown in FIG. 1, and the classifier 4 outputs an inference result based on the input data. The inference result output by the classifier 4 includes a plurality of classes, as shown in the example of FIG. 3. A class is a label classified by the classifier. The classes shown in FIG. 3 indicate labels for estimated values ​​of air resistance coefficients of 0.30 to 0.31, 0.31 to 0.32, 0.32 to 0.33, and 0.33 to 0.34, respectively. The height of the bar graph for each class shown in FIG. 3 indicates the likelihood of each class. The likelihood is the probability that the true value (in the example of FIG. 3, the air resistance coefficient of the actual vehicle) is inferred to be a value included in that class.

[0016] The inference result analysis unit 5 analyzes the inference result output by the classifier 4. The inference result analysis unit 5 includes a maximum likelihood class identification unit 5a, a maximum likelihood class determination unit 5b, and a likelihood difference determination unit 5c. The maximum likelihood class identification unit 5a identifies the class with the highest likelihood among multiple classes included in the inference result as the maximum likelihood class. In the example shown in FIG. 3, the maximum likelihood class identification unit 5a identifies an estimated value of 0.31 to 0.32 as the maximum likelihood class. Next, the maximum likelihood class determination unit 5b determines whether the first predicted data calculated by the prediction model 7 based on the input data input to the input unit 1 is included in the maximum likelihood class. In the example shown in FIG. 3, if the first predicted data P1 calculated by the prediction model 7 is, for example, 0.315 and is included in the class of 0.31 to 0.32, the maximum likelihood class determination unit 5b determines that the first predicted data P1 is included in the maximum likelihood class.

[0017] Furthermore, when the first prediction data P1 is included in the highest likelihood class, the likelihood difference determination unit 5c determines whether the minimum value of the difference between the likelihood of the highest likelihood class among the multiple classes included in the inference result and the likelihood of the other classes is equal to or greater than a predetermined threshold. In the example shown in Fig. 3, the likelihood difference determination unit 5c determines whether the minimum value dT_min of the difference dT between the likelihood of the highest likelihood class (0.31 to 0.32) and the likelihood of the other classes is equal to or greater than a predetermined threshold. Note that this predetermined threshold is a reference value for determining that the likelihood of the highest likelihood class is significantly higher than the likelihood of the other classes.

[0018] If the highest likelihood class determination unit 5b determines that "the first predicted data P1 is not included in the highest likelihood class," the measurement unit 6 calculates second predicted data calculated by CFD simulation or the like without using the prediction model 7 as predicted data, and outputs the calculated data to the output unit 9. In addition, a combination of the input data and the second predicted data calculated by the measurement unit 6 is added to the training dataset D. When a predetermined amount of new data or more is accumulated in the training dataset D, the learning unit 8 updates the first range A1 set by the data range setting unit 2, the parameters of the classifier 4, and the parameters of the prediction model 7 based on the new training dataset D. In addition, even if the highest likelihood class determination unit 5b determines that "the first predicted data P1 is not included in the highest likelihood class," if the likelihood difference determination unit 5c determines that "the minimum value of the difference between the likelihood of the highest likelihood class and the likelihood of the other classes is less than a predetermined threshold," the measurement unit 6 similarly outputs the second predicted data to the output unit 9 as predicted data. On the other hand, if the highest likelihood class determination unit 5b determines that "the first predicted data P1 is included in the highest likelihood class" and the likelihood difference determination unit 5c determines that "the minimum value of the difference between the likelihood of the highest likelihood class and the likelihood of other classes is greater than or equal to a predetermined threshold value," the first predicted data P1 calculated by the prediction model 7 is output to the output unit 9 as predicted data. In addition, without being limited to this embodiment, the information processing device 100 may determine whether the predicted data to be output to the output unit 9 is the first predicted data P1 or the second predicted data P2, based only on whether or not the first predicted data P1 is included in the highest likelihood class.

[0019] Furthermore, the learning unit 8 trains the prediction model 7 based on the training dataset D. The learning unit 8 trains the prediction model 7 in association with the inference result of the classifier 4. That is, when training the prediction model 7, the learning unit 8 compares the predicted data calculated by the prediction model 7 based on the training input data with the inference result output by the classifier 4 based on the training input data, defines a first loss function depending on whether the predicted data is included in the highest likelihood class of the inference result, and updates the parameters of the prediction model 7 based on the first loss function.

[0020] More specifically, the first loss function (L_total) generated by the learning unit 8 has an element (L_classifier) ​​calculated depending on whether the predicted data calculated by the prediction model 7 is included in the highest likelihood class of the inference result. Here, if the predicted data calculated by the prediction model 7 is included in the highest likelihood class of the inference result, L_classifier = 0. On the other hand, if the predicted data calculated by the prediction model 7 is not included in the highest likelihood class of the inference result, L_classifier = 1. Furthermore, the first loss function (L_total) has an element (L1) based on the mean absolute error between the predicted data calculated by the prediction model 7 and the true value, and an element (KL Loss) based on a comparison of the error distribution between the predicted data calculated by the prediction model 7 and the true value with a predetermined required performance distribution. Note that the required performance distribution is a distribution that serves as a criterion for determining whether the error distribution between the predicted data and the true value is a distribution acceptable for the prediction model 7. The required performance distribution may be, for example, a Gaussian distribution with a mean of 0 and a variance of 0.0033.

[0021] The KL Loss is expressed by the following formula (1): In the following formula (1), p is the error distribution between the predicted data calculated by the prediction model 7 and the true value, and q is the required performance distribution.

number

[0022] The first loss function (L_total) is expressed by the following formula (2). That is, the learning unit 8 calculates the first loss function (L_total) by performing a weighted addition of each element of the loss function (L1 Loss, L_classifier, KL Loss). Note that the weighting coefficients α and β in formula (2) differ depending on the types of input data and predicted data to be processed. Furthermore, if the variance of the error distribution between the predicted data and the true value calculated by the prediction model 7 is smaller than the variance of the required performance distribution, the weighting coefficient β of the KL Loss is set to 0. That is, if the variance of the error distribution between the predicted data and the true value calculated by the prediction model 7 is smaller than the variance of the required performance distribution, the learning unit 8 does not use the KL Loss to calculate the first loss function (L_total).

number

[0023] Furthermore, the learning unit 8 simultaneously trains the classifier 4 and the prediction model 7. Specifically, a cross-entropy error depending on whether the predicted data calculated by the measurement unit 6 is included in the most likely class of the inference result output by the classifier 4 is defined as a second loss function. The learning unit 8 then trains the classifier 4 based on the cross-entropy error (second loss function). That is, the learning unit 8 defines a first loss function for training the prediction model 7 and a second loss function for training the classifier depending on whether the predicted data of the prediction model 7 is included in the most likely class of the inference result of the classifier 4, updates the parameters of the prediction model 7 based on the first loss function, and updates the parameters of the classifier 4 based on the second loss function.

[0024] Furthermore, after the classifier 4 and the prediction model 7 have completed learning, the learning unit 8 inputs the learning input data to the prediction model 7 and confirms that the error between the predicted data and the true value is within an acceptable range. The learning unit 8 also inputs the learning input data to the trained classifier 4 and sets the minimum value of the difference between the likelihood of the highest likelihood class in the output inference result and the likelihood of the other classes as the predetermined threshold used by the likelihood difference determination unit 5c for determination.

[0025] The output unit 9 outputs the first predicted data P1 or the second predicted data P2 as predicted data to an output device (not shown). The output device is, for example, a monitor that outputs image information or text information, a speaker or printer that outputs audio information, etc.

[0026] Next, the procedure of the information processing method executed by the information processing device 100 will be described with reference to FIG. First, in step S1, the input data range determination unit 3 of the information processing device 100 determines whether or not the input data input to the input unit 1 is included in the first range A1. If the input data is not included in the first range A1, the information processing device 100 determines that the input data is data that cannot be handled by the prediction model 7, and in step S9, outputs second prediction data P2 calculated using the measurement unit 6 to the output unit 9.

[0027] On the other hand, if the input data is included in the first range A1, in step S2, the input data range determination unit 3 determines whether the input data input to the input unit 1 is included in the second range A2. If the input data is included in the second range A2, the information processing device 100 determines that the input data is data that can be handled by the prediction model 7, and in step S8 outputs first prediction data P1 calculated using the prediction model 7 to the output unit 9.

[0028] On the other hand, if the input data is included in the first range A1 and not included in the second range A2, in step S3, the input data range determination unit 3 determines that the input data input to the input unit 1 is included in the intermediate range Am.

[0029] Next, in step S4, the information processing device 100 inputs the input data to the classifier 4. Further next, in step S5, the highest likelihood class identification unit 5a of the inference result analysis unit 5 identifies the class with the highest likelihood from among the multiple classes included in the inference result output by the classifier 4 as the highest likelihood class.

[0030] Next, in step S6, the highest likelihood class determination unit 5b of the inference result analysis unit 5 determines whether the first predicted data P1 is included in the highest likelihood class. If the first predicted data P1 is not included in the highest likelihood class, the information processing device 100 outputs the second predicted data P2 calculated using the measurement unit 6 to the output unit 9 in step S9.

[0031] On the other hand, if the first predicted data P1 is included in the highest likelihood class, the information processing device 100 determines in step S7 whether the likelihood difference between the highest likelihood class included in the inference result and other classes is equal to or greater than a predetermined threshold. If the likelihood difference between the highest likelihood class and other classes is equal to or greater than the predetermined threshold, the information processing device 100 outputs the first predicted data P1 calculated using the prediction model 7 to the output unit 9 in step S8. On the other hand, if the likelihood difference between the highest likelihood class and other classes is less than the predetermined threshold, the information processing device 100 outputs the second predicted data P2 calculated using the measurement unit 6 to the output unit 9 in step S9.

[0032] As described above, when the input data falls within the intermediate range Am, the information processing device 100 according to this embodiment inputs the input data to the classifier 4, outputs an inference result including multiple classes, and identifies the class with the highest likelihood as the highest likelihood class. The information processing device 100 then determines whether the first predicted data P1 calculated by the prediction model 7 based on the input data falls within the highest likelihood class. If the first predicted data P1 does not fall within the highest likelihood class, the information processing device 100 outputs the second predicted data P2 calculated without using the prediction model 7 as predicted data. That is, when the input data falls within the intermediate range Am, the information processing device 100 determines whether the first predicted data P1 calculated by the prediction model 7 matches the inference result of the classifier 4. If the first predicted data P1 does not match the inference result of the classifier 4, the information processing device 100 determines that the reliability of the first predicted data P1 is lower than the standard required for the performance of the prediction model 7, and outputs the second predicted data calculated without using the prediction model 7 as predicted data. This allows the information processing device 100 to prevent the accuracy of calculation of predicted data from being lower than the required accuracy when the input data is included in the intermediate range Am. Furthermore, since the accuracy of calculation of predicted data can be prevented from being lower than the required accuracy, users other than those who know that "depending on the range of input data, the prediction accuracy of the prediction model 7 may be lower than the required accuracy" can also use the information processing device 100.

[0033] Furthermore, when the first predicted data P1 is included in the highest likelihood class, the information processing device 100 determines whether the minimum value of the difference between the likelihood of the highest likelihood class and the likelihood of the other classes is equal to or greater than a predetermined threshold, and when the minimum value of the difference between the likelihood of the highest likelihood class and the likelihood of the other classes is equal to or greater than the predetermined threshold, outputs the first predicted data P1 as predicted data. As a result, when the likelihood of the highest likelihood class is significantly higher than the likelihood of the other classes, the information processing device 100 determines that the reliability of the inference result of the classifier 4 is equal to or greater than a standard required for the performance of the classifier 4, and when the first predicted data P1 is included in the highest likelihood class, it can output the first predicted data P1 as predicted data.

[0034] Furthermore, when the minimum value of the difference between the likelihood of the highest likelihood class and the likelihood of the other classes is less than a predetermined threshold, the information processing device 100 outputs, as predicted data, second predicted data calculated without using the prediction model 7. As a result, when there is no significant difference between the likelihood of the highest likelihood class and the likelihood of the other classes, the information processing device 100 determines that the reliability of the inference result of the classifier 4 is lower than the standard required for the performance of the classifier 4, and can output the second predicted data P2 as predicted data even if the second predicted data P2 is included in the highest likelihood class. Therefore, the information processing device 100 can prevent the accuracy of calculation of predicted data from being lower than the required accuracy.

[0035] Furthermore, when training the prediction model 7 and the classifier 4, the information processing device 100 compares the predicted data calculated by the prediction model 7 based on the training input data with the inference result output by the classifier 4 based on the training input data, and defines a first loss function for training the prediction model 7 and a second loss function for training the classifier 4 depending on whether the predicted data is included in the highest likelihood class of the inference result. In this way, the information processing device 100 can train the prediction model 7 and the classifier 4 in cooperation with each other so that the first predicted data P1 calculated by the prediction model 7 matches the inference result of the classifier 4.

[0036] Furthermore, the information processing device 100 calculates the second predicted data P2 using a simulation. As a result, even if the information processing device 100 determines that the reliability of the first predicted data P1 is lower than the standard required for the performance of the prediction model 7, the information processing device 100 can output the second predicted data P2 calculated with accuracy equal to or higher than the required standard as predicted data.

[0037] The input data is data relating to the shape of the vehicle, and the predicted data is data indicating the aerodynamic performance of the vehicle. This allows the information processing device 100 to output predicted data indicating the aerodynamic performance of the vehicle with accuracy that meets or exceeds the required standard, based on the data relating to the shape of the vehicle.

[0038] Furthermore, without being limited to this embodiment, the input data may be time-series data regarding a person's posture, and the predicted data may be data quantitatively indicating the person's fatigue level or proficiency in a predetermined task. Specifically, the input unit 1 of the information processing device 100 extracts features indicating changes in the person's posture from a video or a series of images captured by an imaging device, and the information processing device 100 calculates the person's fatigue level or proficiency in a predetermined task as a numerical value, for example, expressed on a five-point scale, based on the extracted features. Furthermore, if the first predicted data P1 is not included in the highest likelihood class, the measurement unit 6 calculates the second predicted data P2 based on a rule base. As a result, the information processing device 100 can output predicted data indicating the person's fatigue level or proficiency in a predetermined task with accuracy above the required standard, based on the time-series data regarding the person's posture.

[0039] Alternatively, the input data may be image data of a person, and the predicted data may be biometric data indicating both or either of the person's height and weight. Specifically, the input unit 1 of the information processing device 100 may extract features from an image captured by an imaging device, and the information processing device 100 may calculate the person's height and / or weight based on the extracted features. If the first predicted data P1 is not included in the highest likelihood class, the measurement unit 6 calculates the second predicted data P2 based on a rule base. This allows the information processing device 100 to output predicted data indicating both or either of the person's height and weight with accuracy above the required standard based on the image data of the person. Furthermore, the predicted data is not limited to the person's height or weight, and may be biometric information such as body temperature, for example.

[0040] The input data may be still or video image data of the driver while driving, and the prediction data may be data that quantitatively indicates the driver's concentration level or fatigue level. [Explanation of symbols]

[0041] 100...Information processing device 2a...First range calculation unit 2b...Second range calculation section 2c...Intermediate range calculation section 3...Input data range determination section 4...Classifier 5a...Highest likelihood class identification part 5b...Highest likelihood class determination section 6...Measuring section 7…Prediction model 9...Output section A1...First range A2...Second range Am: Mid-range D...Training dataset

Claims

1. 1. An information processing method for outputting prediction data based on input data using an information processing device having a prediction model trained using a training dataset including a plurality of pieces of training input data, comprising: The information processing device includes: calculating a first range that is a data range including the plurality of learning input data; calculating a second range, which is a data range narrower than the first range and includes the input data when an error between the predicted data calculated by the prediction model and a true value is equal to or smaller than a predetermined value; calculating an intermediate range that is included in the first range and not included in the second range; determining whether the input data input to the information processing device is within the intermediate range; If the input data is within the intermediate range, inputting the input data to a classifier and outputting an inference result including a plurality of classes; Identifying the class with the highest likelihood as a most likely class among the plurality of classes; determining whether first predicted data calculated by the prediction model based on the input data is included in the highest likelihood class; When the first predicted data is not included in the highest likelihood class, second predicted data calculated without using the prediction model is output as the predicted data.

2. The information processing device includes: If the first prediction data is included in the highest likelihood class, it is determined whether or not a minimum value of a difference between the likelihood of the highest likelihood class and the likelihood of other classes among a plurality of classes included in the inference result is equal to or greater than a predetermined threshold value; The information processing method according to claim 1 , wherein when a minimum value of the difference between the likelihood of the highest likelihood class and the likelihood of the other classes is equal to or greater than a predetermined threshold, the first predicted data is output as the predicted data.

3. The information processing device includes: The information processing method according to claim 2 , wherein when a minimum value of the difference between the likelihood of the highest likelihood class and the likelihood of the other classes is less than a predetermined threshold, the second predicted data is output as the predicted data.

4. The information processing device includes: When training the prediction model and the classifier, the prediction data calculated by the prediction model based on the training input data is compared with the inference result output by the classifier based on the training input data; defining a first loss function for training the prediction model and a second loss function for training the classifier depending on whether the prediction data is included in the highest likelihood class of the inference result; updating parameters of the prediction model based on the first loss function; The information processing method according to claim 1 , further comprising updating parameters of the classifier based on the second loss function.

5. 5. The information processing method according to claim 1, wherein the information processing device calculates the second predicted data using a simulation.

6. the input data is data relating to the shape of a vehicle, 6. The information processing method according to claim 1, wherein the prediction data is data indicating aerodynamic performance of the vehicle.

7. the input data is time-series data relating to a person's posture, 6. The information processing method according to claim 1, wherein the prediction data is data indicating the person's level of fatigue or their level of proficiency in a predetermined task.

8. the input data is image data of a person; 6. The information processing method according to claim 1, wherein the prediction data is data indicating both or either one of a height and a weight of the person.

9. An information processing device having a prediction model trained using a training dataset including a plurality of pieces of training input data, and outputting prediction data based on the input data, a first range calculation unit that calculates a first range, which is a data range including the plurality of learning input data; a second range calculation unit that calculates a second range, which is a data range narrower than the first range and includes the input data when an error between the predicted data calculated by the prediction model and a true value is equal to or smaller than a predetermined value; an intermediate range calculation unit that calculates an intermediate range that is included in the first range and not included in the second range; an input data range determination unit that determines whether the input data input to the information processing device is included in the intermediate range; a classifier that outputs an inference result including a plurality of classes based on the input data when the input data is within the intermediate range; a highest likelihood class identification unit that identifies a class with the highest likelihood as a highest likelihood class among the plurality of classes; a maximum likelihood class determination unit that determines whether first predicted data calculated by the prediction model based on the input data is included in the maximum likelihood class; a measurement unit that calculates second predicted data without using the prediction model when the first predicted data is not included in the highest likelihood class; an output unit that outputs, when the first predicted data is not included in the highest likelihood class, second predicted data calculated without using the prediction model as the predicted data.

Images