Information processing device, program, and information processing method

Abstract

An information processing device (100) is provided with: an attention mechanism unit (113) that calculates a context variable by weighting and adding a plurality of variables, which are time series, using an attention mechanism training model, which is an attention mechanism training model; a determination unit (114) that estimates one determination included in a plurality of determinations on the basis of the reliability of the plurality of determinations calculated on the basis of the context variable and the latest one variable included in the plurality of variables; a storage unit (101) that stores result information in which the context variable is associated with the one determination; and an evaluation unit (115) that evaluates at least the training state of the attention mechanism training model on the basis of the result information.

Description

Technical Field

[0001] The present disclosure relates to an information processing device, a program, and an information processing method. Background Art

[0002] As a technique for improving the estimation accuracy of a training model, there is an attention mechanism. For example, Non-Patent Document 1 describes that the accuracy of translation can be improved by using the attention mechanism in natural language translation based on a neural network.

[0003] Prior art literature

[0004] Non-patent literature

[0005] Non-patent literature 1: Minh-Thang Luong et al., "Effective Approaches to Attention-based Neural Machine Translation", arXiv preprent arXiv:1508.04025, 18Aug 2015 Summary of the invention

[0006] Problem that the invention aims to solve

[0007] However, the internal processing of the training model using deep reinforcement training is a black box and cannot be seen. Therefore, the user cannot easily determine whether the training model is effectively trained.

[0008] Therefore, an object of one or more schemes of the present disclosure is to easily grasp the training status of a training model using an attention mechanism.

[0009] Means used to solve problems

[0010] An information processing device of one scheme of the present disclosure is characterized in that it comprises: an attention mechanism unit, which uses a training model of the attention mechanism, i.e., the attention mechanism training model, to weight and add multiple variables as time series to calculate a context variable; a judgment unit, which estimates a judgment included in the multiple judgments based on the reliability of multiple judgments calculated based on the context variable and the latest variable included in the multiple variables; a storage unit, which stores result information obtained by corresponding the context variable to the one judgment; and an evaluation unit, which evaluates at least the training status of the attention mechanism training model based on the result information.

[0011] A program of one scheme of the present disclosure is characterized in that the program enables a computer to function as an attention mechanism unit, a judgment unit, a storage unit and an evaluation unit, the attention mechanism unit uses a training model of the attention mechanism, i.e., the attention mechanism training model, to weight and add multiple variables as time series to calculate a context variable, the judgment unit estimates a judgment included in the multiple judgments based on the reliability of multiple judgments calculated based on the context variable and the latest variable included in the multiple variables, the storage unit stores result information obtained by corresponding the context variable to the one judgment, and the evaluation unit evaluates at least the training status of the attention mechanism training model based on the result information.

[0012] An information processing method of one scheme of the present disclosure is characterized in that a training model of an attention mechanism, i.e., an attention mechanism training model, is used to weight and add a plurality of variables as a time series to calculate a context variable, and based on the reliability of a plurality of judgments calculated based on the context variable and a latest variable included in the plurality of variables, a judgment included in the plurality of judgments is estimated, result information obtained by corresponding the context variable to the one judgment is stored, and based on the result information, at least the training status of the attention mechanism training model is evaluated.

[0013] Effects of the Invention

[0014] According to one or more aspects of the present disclosure, it is possible to easily grasp the training status of a training model using an attention mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a block diagram schematically showing the configuration of the information processing device according to the first embodiment.

[0016] Figure 2 (A) and (B) are block diagrams showing hardware configuration examples.

[0017] Figure 3 This is a schematic diagram for explaining the processing in the information processing device according to the first embodiment.

[0018] Figure 4 This is a block diagram schematically showing the structure of an information processing device according to the second embodiment.

[0019] Figure 5 This is a schematic diagram for explaining the processing in the information processing device according to the second embodiment.

[0020] Figure 6 This is a block diagram schematically showing the structure of an information processing device according to a third embodiment.

[0021] Figure 7This is a schematic diagram for explaining the processing in the information processing device according to the third embodiment.

[0022] Figure 8 This is a block diagram schematically showing the structure of an information processing device according to a fourth embodiment.

[0023] Fig. 9 This is a schematic diagram for explaining the processing in the information processing device according to the fourth embodiment. DETAILED DESCRIPTION

[0024] Implementation method 1.

[0025] Figure 1 This is a block diagram schematically showing the configuration of information processing device 100 according to Embodiment 1.

[0026] The information processing device 100 includes a storage unit 101 , a communication unit 102 , an input unit 103 , a display unit 104 , and a control unit 110 .

[0027] The storage unit 101 stores programs and data necessary for processing in the information processing device 100 .

[0028] For example, the storage unit 101 stores at least an attention mechanism training model, which is a training model used in the attention mechanism executed by the control unit 110. In addition, in the first embodiment, as described later, the storage unit 101 also stores an extraction training model and a judgment training model.

[0029] In addition, the storage unit 101 stores result information in which a judgment result determined by the control unit 110 using the estimation result based on the attention mechanism is associated with the estimation result.

[0030] The communication unit 102 communicates with other devices. For example, the communication unit 102 communicates with other devices using a network such as the Internet.

[0031] The input unit 103 receives input from a user of the information processing device 100 .

[0032] The display unit 104 displays information to a user of the information processing device 100. For example, the display unit 104 displays various screen images.

[0033] The control unit 110 controls the processing in the information processing device 100. For example, the control unit 110 adds weights to variables required for judgment, i.e., state variables, through an attention mechanism, thereby calculating context state variables and estimating a certain judgment based on the context state variables. Then, the control unit 110 associates the context state variables with the judgment estimated based on the context state variables, and stores the result information in the storage unit 101.

[0034] In addition, below, state variables are also referred to as variables for short, and context state variables are also referred to as context variables for short.

[0035] Furthermore, the control unit 110 evaluates the training status of the training model used at least in the attention mechanism using the result information stored in the storage unit 101. In addition, in embodiment 1, as described later, the control unit 110 performs the evaluation of the training status of the extraction training model, the attention mechanism training model, and the judgment training model.

[0036] The control unit 110 includes a data acquisition unit 111 , a variable extraction unit 112 , an attention mechanism unit 113 , a determination unit 114 , and an evaluation unit 115 .

[0037] The data acquisition unit 111 acquires input data. The data acquisition unit 111 may acquire input data, for example, via the communication unit 102. In addition, when the input data is stored in the storage unit 101, the data acquisition unit 111 may acquire the input data from the storage unit 101.

[0038] The variable extraction unit 112 extracts state variables, which are variables that can be used for determination, from the input data acquired by the data acquisition unit 111 .

[0039] Here, the variable extraction unit 112 extracts the state variables using an extraction training model, which is a training model for extracting the state variables from the input data.

[0040] The attention mechanism unit 113 calculates the context state variable by performing a weighted sum based on a well-known attention mechanism on the state variables extracted by the variable extraction unit 112. For example, the attention mechanism unit 113 uses the training model stored in the storage unit 101 to weight the state variables extracted by the variable extraction unit 112, and adds the weighted state variables, thereby calculating the context state variable as the estimation result.

[0041] The judgment unit 114 estimates one judgment included in the plurality of judgments based on the reliability of the plurality of judgments calculated based on the context state variable estimated by the attention mechanism unit 113 and the latest state variable included in the plurality of state variables. Then, the judgment unit 114 associates the one judgment with the context state variable and stores it in the storage unit 101 as result information.

[0042] Here, the judgment unit 114 performs the estimation using a judgment training model which is a training model for estimating one judgment based on context variables.

[0043] The evaluation unit 115 evaluates the training status of at least the training model used by the attention mechanism unit 113, that is, the attention mechanism training model, based on the result information stored in the storage unit 101.

[0044] In addition, in Embodiment 1, the evaluation unit 115 evaluates the training status of the extraction training model, the attention mechanism training model, and the judgment training model. However, the evaluation unit 115 evaluates the training status of the attention mechanism training model and the judgment training model without extracting state variables from the input data.

[0045] For example, the evaluation unit 115 performs clustering according to each of the plurality of determinations, thereby determining a plurality of clusters, and performs evaluation based on the distance or similarity between the plurality of clusters. In this case, the shorter the distance or the higher the similarity, the lower the evaluation.

[0046] For example, part or all of the control unit 110 described above may be Figure 2 As shown in (A), the information processing device 100 can be composed of a memory 10 and a processor 11 such as a CPU (Central Processing Unit) that executes the program stored in the memory 10. In other words, the information processing device 100 can be implemented by a so-called computer. Such a program can be provided through a network, and can also be provided by recording in a recording medium. That is, such a program can also be provided as a program product, for example.

[0047] In addition, part or all of the control unit 110, for example, Figure 2 As shown in (B), the processing circuit 12 may be composed of a single circuit, a composite circuit, a processor operated by a program, a parallel processor operated by a program, an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

[0048] As described above, the control unit 110 can be implemented by a processing circuit network.

[0049] The storage unit 101 can be realized by a storage device such as a HDD (Hard Disk Drive) or an SSD (Solid state Drive).

[0050] The communication unit 102 can be realized by a communication interface such as a NIC (Network Interface Card).

[0051] The input unit 103 can be realized by an input interface such as a keyboard or a mouse.

[0052] The display unit 104 can be realized by a display.

[0053] Figure 3 This is a schematic diagram for explaining the processing in the information processing device 100 according to the first embodiment.

[0054] First, the data acquisition unit 111 acquires input data X t-n , X t-n+1 , X t-1 , X t (S10). Here, input data X t-n , X t-n+1 , X t-1 , X t tn is a sensor value as an observation value, and is assumed to be data of a time series of tn, t-n+1, ​​t-1, t (t and n are positive integers). For example, image data can be used as input data.

[0055] The data acquisition unit 111 acquires the input data X t-n , X t-n+1 , X t-1 , X t Provided to the variable extraction unit 112.

[0056] The variable extraction unit 112 extracts the input data X t-n , X t-n+1 , X t-1 , X t The state variable S is extracted from the equation as a variable useful for the determination by the determination unit 114. t-n , S t-n+1 , S t-1 , S t (S11).

[0057] Here, the variable extraction unit 112 uses the extraction training model as the neural network model stored in the storage unit 101 to extract the input data X t-n , X t-n+1 , X t-1 , X t Extract the state variable S t-n , S t-n+1 , S t-1 , S t .

[0058] The variable extraction unit 112 extracts the state variable S t-n , S t-n+1 , S t-1 , S t Provided to the attention mechanism unit 113.

[0059] In addition, here, the variable extraction unit 112 uses the extraction training model, but the first embodiment is not limited to such an example, and some functions are used to extract the state variable S t-n , S t-n+1 , S t-1 , S t That's it.

[0060] The attention mechanism unit 113 is for the state variable S t-n , S t-n+1 , S t-1 , S t , use the training model to estimate the weight value, calculate the weighted sum, and thus calculate the context state variable (S12).

[0061] The attention mechanism unit 113 provides the calculated context state variable to the judgment unit 114.

[0062] The determination unit 114 makes a determination based on the context state variable and the latest state variable St ( S13 ).

[0063] Here, the judgment unit 114 uses the judgment training model as a neural network model stored in the storage unit 101 to estimate the judgment based on the context state variables and the latest state variables.

[0064] Then, the determination unit 114 associates the determination with the context state variable and stores the result information in the storage unit 101 as result information, thereby accumulating the result information ( S14 ).

[0065] The evaluation unit 115 uses the result information stored in the storage unit 101 to evaluate the training status of at least the training model used by the attention mechanism unit 113.

[0066] For example, the evaluation unit 115 converts the N-dimensional data obtained by clustering the result information according to each judgment into data of a lower dimension for easy evaluation (S15). Specifically, the evaluation unit 115 uses T-sne (T-Eistributed Stochastic Neighbor Embedding) to convert the N-dimensional data into two-dimensional data, thereby visualizing the clustering of each judgment.

[0067] Then, the evaluation unit 115 calculates, for example, the distance or similarity between clusters as an evaluation value, thereby evaluating the training status ( S16 ).

[0068] For example, the evaluation unit 115 performs evaluation by comparing the evaluation value between clusters with a threshold value. Specifically, the evaluation unit 115 determines that the training is insufficient when the distance between clusters is smaller than a predetermined threshold value or when the similarity between clusters is higher than a predetermined threshold value.

[0069] In addition, the determination result of the evaluation unit 115 may be displayed on the display unit 104 , for example.

[0070] As described above, according to implementation mode 1, the training status of the training model using the attention mechanism can be easily grasped.

[0071] Implementation method 2.

[0072] Figure 4 This is a block diagram schematically showing the configuration of information processing device 200 according to Embodiment 2.

[0073] The information processing device 200 includes a storage unit 101 , a communication unit 102 , an input unit 103 , a display unit 104 , and a control unit 210 .

[0074] The storage unit 101 , the communication unit 102 , the input unit 103 , and the display unit 104 of the information processing device 200 of the second embodiment are the same as those of the information processing device 100 of the first embodiment.

[0075] The control unit 210 controls processing in the information processing device 200 .

[0076] The control unit 210 in the second embodiment performs the following processing in addition to the same processing as the control unit 110 in the first embodiment.

[0077] The control unit 210 performs training of the training model using the additional training data, ie, the additional training data, based on the evaluation result of the training state.

[0078] The control unit 210 includes a data acquisition unit 111 , a variable extraction unit 112 , an attention mechanism unit 113 , a determination unit 114 , an evaluation unit 215 , and an additional training unit 216 .

[0079] The data acquisition unit 111, variable extraction unit 112, attention mechanism unit 113 and judgment unit 114 of the control unit 210 in implementation mode 2 are the same as the data acquisition unit 111, variable extraction unit 112, attention mechanism unit 113 and judgment unit 114 of the control unit 110 in implementation mode 1.

[0080] The evaluation unit 215 uses the result information stored in the storage unit 101 to evaluate the training status of at least the training model used by the attention mechanism unit 113.

[0081] Then, the evaluation unit 215 provides the evaluation result to the additional training unit 216. For example, the evaluation unit 215 compares the evaluation value with a threshold value for each combination of two clusters to generate evaluation information indicating whether the training is sufficient, and provides the evaluation information to the additional training unit 216.

[0082] The additional training unit 216 refers to the evaluation information from the evaluation unit 215 and provides the additional training data to the variable extraction unit 112 , thereby performing additional training.

[0083] Here, the additional training unit 216 uses additional training data, i.e., additional training data, to train at least the attention mechanism training model when the evaluation of the evaluation unit 215 is lower than a predetermined threshold. In embodiment 2, the additional training unit 216 trains the extraction training model, the judgment training model, and the attention mechanism training model.

[0084] For example, the additional training unit 216 uses the training data judged as correct answers in the plurality of judgments with an evaluation lower than a predetermined threshold as additional training data to perform training. In other words, the additional training unit 216 provides the training data classified into the two clusters judged as insufficiently trained to the variable extraction unit 112 as additional training data. Here, the additional training data can be obtained from other devices via the communication unit 102, for example, or stored in the storage unit 101. As to where to obtain the additional training data, for example, the user can also indicate via the input unit 103.

[0085] Figure 5 This is a schematic diagram for explaining the processing in the information processing device 200 according to the second embodiment.

[0086] Figure 5 The processing of S10~S15 and Figure 3 The processes of S10 to S15 shown are the same.

[0087] In the second embodiment, the evaluation unit 215 evaluates the training status by calculating the distance or similarity between clusters as an evaluation value, and generates evaluation information (S26) indicating the evaluation result. The evaluation information indicates whether the training is sufficient according to the combination of each two clusters. The generated evaluation information is provided to the additional training unit 216.

[0088] The additional training unit 216 refers to the evaluation information, generates training data classified into the cluster determined to be insufficiently trained as additional training data ( S27 ), and provides the additional training data to the variable extraction unit 112 , thereby performing additional training.

[0089] As described above, according to Embodiment 2, in a training model using an attention mechanism, it is possible to additionally train insufficiently trained clusters.

[0090] Here, the evaluation unit 215 may use a single threshold to determine whether the training is sufficient, but, for example, risk management of the determination can be performed by using multiple thresholds. Specifically, for the determinations that cannot be wrong, such as "stopping" and "accelerating" the vehicle, the distance between clusters must be long or the similarity must be low, and therefore, risk management of the determination can be performed by adjusting the threshold.

[0091] Implementation method 3.

[0092] Figure 6 This is a block diagram schematically showing the configuration of information processing device 300 according to Embodiment 3.

[0093] The information processing device 300 includes a storage unit 101 , a communication unit 102 , an input unit 103 , a display unit 104 , and a control unit 310 .

[0094] The storage unit 101 , the communication unit 102 , the input unit 103 , and the display unit 104 of the information processing device 300 of the third embodiment are the same as those of the information processing device 100 of the first embodiment.

[0095] The control unit 310 controls processing in the information processing device 300 .

[0096] The control unit 310 in the third embodiment performs the following processing in addition to the same processing as the control unit 110 in the first embodiment.

[0097] The control unit 310 selects training data based on the evaluation result of the training state, and trains the training model using the selected training data.

[0098] The control unit 310 includes a data acquisition unit 111 , a variable extraction unit 112 , an attention mechanism unit 113 , a judgment unit 114 , an evaluation unit 315 , a training data selection unit 317 , and a training unit 318 .

[0099] The data acquisition unit 111, variable extraction unit 112, attention mechanism unit 113 and judgment unit 114 of the control unit 310 in implementation mode 3 are the same as the data acquisition unit 111, variable extraction unit 112, attention mechanism unit 113 and judgment unit 114 of the control unit 110 in implementation mode 1.

[0100] The evaluation unit 315 uses the result information stored in the storage unit 101 in the same manner as in implementation mode 1 to evaluate the training status of at least the training model used by the attention mechanism unit 113.

[0101] In the third embodiment, the evaluation unit 315 supplies the training data selection unit 317 with evaluation value information indicating the evaluation value for each combination of two clusters.

[0102] The training data selection unit 317 refers to the evaluation value information from the evaluation unit 315 and selects training data for at least training the attention mechanism training model.

[0103] Here, the training data selection unit 317 selects training data in such a manner that the lower the evaluation corresponding to a judgment, the more training data that takes the judgment as a correct answer. In other words, the training data selection unit 317 selects training data in such a manner that the lower the evaluation based on the evaluation value indicated by the evaluation value information, in other words, the shorter the distance or the higher the similarity, the more training data that is classified into the cluster. In addition, the training data may be stored in the storage unit 101 or in other devices. In the case where the training data is stored in other devices, the training data selection unit 317 may access the other devices via the communication unit 102 to select the training data.

[0104] The training unit 318 uses the training data selected by the training data selection unit 317 to train at least the attention mechanism training model.

[0105] For example, the training unit 318 performs training by providing the training data selected by the training data selection unit 317 to the variable extraction unit 112 .

[0106] Figure 7 This is a schematic diagram for explaining the processing in the information processing device 300 according to the third embodiment.

[0107] Figure 7 The following is a process for performing training using training data in the information processing device 300 .

[0108] In addition, as a premise, the training data selection unit 317 provides the training data selected without referring to the evaluation value information as initial training data to the training unit 318. The training unit 318 performs initial training by providing the initial training data to the variable extraction unit 112, and selects training data based on the evaluation results in the initial training.

[0109] Figure 7 The processing of S11 to S15 and Figure 3 The processes of S11 to S15 shown are the same.

[0110] In the third embodiment, the evaluation unit 315 evaluates the training state by calculating the distance or similarity between clusters as an evaluation value, and generates evaluation value information indicating the evaluation value for each combination of two clusters (S36). The generated evaluation value information is provided to the training data selection unit 317.

[0111] The training data selection unit 317 refers to the evaluation value information and selects training data such that the lower the evaluation value indicated by the evaluation value information, the greater the number of training data classified into the cluster ( S37 ). The training data selection unit 317 then provides the selected training data to the training unit 318 .

[0112] The training unit 318 performs training by providing the training data selected by the training data selection unit 317 to the variable extraction unit 112 ( S38 ).

[0113] As described above, according to Embodiment 3, when training a training model using an attention mechanism, training can be performed efficiently by selecting training data that should be focused on.

[0114] In addition, the training data selection unit 317 selects training data in such a manner that the lower the evaluation value indicated by the evaluation value information, the more training data is classified into the cluster, but the third embodiment is not limited to such an example. For example, for clusters of judgments that cannot be wrong, such as "stopping" and "accelerating" the vehicle, the clusters are pre-set in the training data selection unit 317 as clusters that should be trained intensively in advance, thereby enabling the training data selection unit 317 to select such clusters in such a manner that the training data of such clusters increases. Specifically, the training data selection unit 317 adds or multiplies weight values ​​that decrease the evaluation value for the clusters that should be trained intensively, thereby increasing the number of selected training data. For example, such a setting may also be made by the user via the input unit 103.

[0115] Implementation method 4.

[0116] Figure 8 This is a block diagram schematically showing the structure of information processing device 400 according to Embodiment 4.

[0117] The information processing device 400 includes a storage unit 101 , a communication unit 102 , an input unit 103 , a display unit 104 , and a control unit 410 .

[0118] The storage unit 101 , the communication unit 102 , the input unit 103 , and the display unit 104 of the information processing device 400 of the fourth embodiment are the same as those of the information processing device 100 of the first embodiment.

[0119] The control unit 410 controls the processing in the information processing device 400 .

[0120] The control unit 410 in the fourth embodiment performs the following processing in addition to the same processing as the control unit 110 in the first embodiment.

[0121] The control unit 410 determines whether to continue the training based on the evaluation result of the training state. If it is determined that the training should be continued, the control unit 410 continues the training, and if it is determined that the training should not be continued, the control unit 410 ends the training.

[0122] The control unit 410 includes a data acquisition unit 111 , a variable extraction unit 112 , an attention mechanism unit 113 , a determination unit 114 , an evaluation unit 215 , a training unit 418 , and a training continuation determination unit 419 .

[0123] The data acquisition unit 111, variable extraction unit 112, attention mechanism unit 113 and judgment unit 114 of the control unit 410 in implementation mode 4 are the same as the data acquisition unit 111, variable extraction unit 112, attention mechanism unit 113 and judgment unit 114 of the control unit 110 in implementation mode 1.

[0124] The evaluation unit 215 in the fourth embodiment is the same as the evaluation unit 215 in the second embodiment. However, in the fourth embodiment, the evaluation unit 215 provides the evaluation information to the training continuation determination unit 419 .

[0125] The training continuation determination unit 419 refers to the evaluation information from the evaluation unit 215 to determine whether to at least continue the training of the attention mechanism training model.

[0126] For example, the training continuation determination unit 419 determines to continue training when all or part of the evaluations based on the evaluation values ​​indicated by the evaluation information are lower than a predetermined threshold, in other words, when the distance is shorter than a predetermined threshold or the similarity is higher than a predetermined threshold.

[0127] In addition, some of the evaluations may be evaluations of a predetermined number or evaluations of predetermined clusters. For example, when all evaluations of important clusters that cannot allow errors are greater than a threshold, the training continuation determination unit 419 may determine not to continue training.

[0128] When the training continuation determination unit 419 determines that the training is to be continued, the training unit 418 performs training by providing the training data to the variable extraction unit 112. On the other hand, when the training continuation determination unit 419 determines that the training is not to be continued, the training unit 418 does not provide the training data to the variable extraction unit 112 and ends the training.

[0129] The training data may be stored in the storage unit 101 or in another device. If the training data is stored in another device, the training unit 418 may access the other device via the communication unit 102 to obtain the training data.

[0130] Fig. 9 This is a schematic diagram for explaining the processing in information processing device 400 according to the fourth embodiment.

[0131] Fig. 9 The following is a process for performing training using training data in the information processing device 400 .

[0132] Furthermore, as a premise, the training unit 418 provides the training data as initial training data to the variable extraction unit 112 to perform initial training, and determines whether to continue the training based on the evaluation result of the initial training.

[0133] Fig. 9 The processing of S11 to S15 and Figure 3 The processes of S11 to S15 shown are the same.

[0134] In the fourth embodiment, the evaluation unit 215 evaluates the training state by calculating the distance or similarity between clusters as an evaluation value, and generates evaluation information (S46) indicating the evaluation result. The evaluation information indicates whether the training is sufficient according to the combination of each two clusters. The generated evaluation information is provided to the training continuation judgment unit 419.

[0135] The training continuation determination unit 419 refers to the evaluation information from the evaluation unit 215 to determine whether to continue the training ( S47 ).

[0136] Then, when the training continuation determination unit 419 determines that the training is to be continued, the training unit 418 performs the training by providing the training data to the variable extraction unit 112 ( S48 ).

[0137] As described above, according to Embodiment 4, when training a training model using an attention mechanism, the training can be terminated when the training is sufficiently performed. Therefore, the training can be performed efficiently.

[0138] In addition, similar to Embodiment 2, the evaluation unit 215 may use a single threshold value to determine whether the training is sufficient, but for example, it is possible to manage the risk of judgment by using multiple threshold values. Specifically, for judgments that cannot be wrong, such as "stopping" and "accelerating" the vehicle, the distance between clusters must be long or the similarity must be small, so it is possible to manage the risk of judgment by adjusting the threshold value.

[0139] Description of Reference Numerals

[0140] 100, 200, 300, 400 information processing device, 101 storage unit, 102 communication unit, 103 input unit, 104 display unit, 110, 210, 310, 410 control unit, 111 data acquisition unit, 112 variable extraction unit, 113 attention mechanism unit, 114 judgment unit, 115, 215, 315 evaluation unit, 216 additional training unit, 317 training data selection unit, 318, 418 training unit, 419 training continuation judgment unit.

Claims

1. An information processing device, characterized in that: The information processing device comprises: An attention mechanism unit, which uses an attention mechanism training model, i.e., an attention mechanism training model, to weight and add a plurality of variables as a time series, thereby calculating a context variable; a judgment unit that estimates one judgment included in the plurality of judgments based on the reliability of the plurality of judgments calculated based on the context variable and a latest one of the plurality of variables; a storage unit for storing result information obtained by associating the context variable with the one judgment; as well as An evaluation unit, which evaluates at least the training status of the attention mechanism training model based on the result information.

2. The information processing device according to claim 1, characterized in that The judgment unit estimates the one judgment using a judgment training model, where the judgment training model is a training model for estimating the one judgment based on the context variable. The evaluation unit performs the evaluation of the judgment training model and the attention mechanism training model.

3. The information processing device according to claim 2, characterized in that: The information processing device further includes a variable extraction unit that extracts the variable from input data.

4. The information processing device according to claim 3, characterized in that: The variable extraction unit extracts the variable using an extraction training model, which is a training model for extracting the variable from the input data. The evaluation unit performs the evaluation of the extraction training model, the judgment training model and the attention mechanism training model.

5. The information processing device according to claim 1, characterized in that The information processing device further includes a variable extraction unit that extracts the variable from input data.

6. The information processing device according to claim 5, characterized in that: The variable extraction unit extracts the variable using an extraction training model, which is a training model for extracting the variable from the input data. The evaluation unit performs the evaluation of the extraction training model and the attention mechanism training model.

7. The information processing device according to any one of claims 1 to 6, characterized in that: The evaluation unit determines a plurality of clusters by clustering according to each of the plurality of determinations, and performs the evaluation based on distances or similarities between the plurality of clusters.

8. The information processing device according to any one of claims 1 to 7, characterized in that: The information processing device further includes an additional training unit that uses additional training data, namely, additional training data, to train at least the attention mechanism training model when the evaluation is lower than a predetermined threshold.

9. The information processing device according to claim 8, characterized in that: The additional training unit uses, as the additional training data, training data for which a judgment in which the evaluation is lower than the predetermined threshold value among the plurality of judgments is a correct answer.

10. The information processing device according to any one of claims 1 to 7, characterized in that: The information processing device further comprises: a training data selection unit, which selects training data for training at least the attention mechanism training model according to the evaluation; and A training unit, which uses the selected training data to at least train the attention mechanism training model.

11. The information processing device according to claim 10, characterized in that: The training data selection unit performs the selection so that the lower the evaluation of the one judgment is, the greater the number of the training data that take the one judgment as a correct answer is.

12. The information processing device according to any one of claims 1 to 7, characterized in that: The information processing device further comprises: a training continuation determination unit, which determines whether to at least continue the training of the attention mechanism training model based on the evaluation; and A training unit, which, when it is determined that the training should be continued, continues the training using training data for training at least the attention mechanism training model, and ends the training when it is determined that the training should not be continued.

13. The information processing device according to claim 12, characterized in that: The training continuation determination unit determines that the training should be continued when the evaluations in all or part of the plurality of determinations are lower than a predetermined threshold value.

14. A program, characterized in that The program enables the computer to function as an attention mechanism unit, a judgment unit, a storage unit, and an evaluation unit. The attention mechanism unit uses an attention mechanism training model, namely an attention mechanism training model, to weight and add a plurality of variables as a time series, thereby calculating a context variable. The judgment unit estimates one judgment included in the plurality of judgments based on the reliability of the plurality of judgments calculated based on the context variable and a latest variable included in the plurality of variables, The storage unit stores result information obtained by associating the context variable with the one judgment. The evaluation unit at least evaluates the training status of the attention mechanism training model based on the result information.

15. An information processing method, characterized in that: The training model using the attention mechanism, i.e., the attention mechanism training model, weights and adds multiple variables as time series to calculate the context variable, estimating a judgment included in the plurality of judgments based on the reliability of the plurality of judgments calculated based on the context variable and a latest variable included in the plurality of variables, storing result information obtained by matching the context variable with the one judgment, Based on the result information, at least the training status of the attention mechanism training model is evaluated.

Images