Information processing device and information processing method

Abstract

Provided are an information processing device, an information processing method, and a program that optimize a data set while suppressing calculation costs. This information processing device includes a memory that stores a program and a processor that executes processing according to the program, wherein the processor: executes the program to generate an updated data set by adding data to or deleting data from a reference data set according to a policy; converts the updated data set into converted data, which is data for entering the updated data set into a regression model that estimates the accuracy with which a discrimination model discriminates the updated data set; and determines a policy on the basis of an evaluation value of an objective function including the discrimination accuracy estimated by the regression model.

Description

Information processing device and information processing method

[0001] The present disclosure relates to techniques for optimizing datasets used in machine learning.

[0002] The development of AI (Artificial Intelligence) requires the preparation of datasets to be used in machine learning. A dataset is a group of data used for training, validation, and inference of machine learning models. Commercially available datasets are versatile, but contain a huge amount of data and are not suitable for individual machine learning. It is desirable to prepare an appropriate dataset, neither too much nor too little, as training data for the desired machine learning. Therefore, it is necessary to expand or delete data from the original dataset.

[0003] Data augmentation is the process of modifying some of the data in a dataset and adding the modified data to the data set. The purpose of data augmentation is to improve the accuracy of the model. Data reduction is the process of deleting some of the data in a dataset to reduce the number of data in the dataset. The purpose of data reduction is to reduce the computational cost required for training.

[0004] An automatic data augmentation technique is known that automatically performs data expansion and reduction without manual work for a data set of a discriminant model that performs class classification (see, for example, Non-Patent Document 1). The automatic data augmentation technique disclosed in Non-Patent Document 1 automatically performs data expansion so as to maximize the discrimination accuracy of class classification.

[0005] Ekin D. Cubuk, and 4 others, “Auto Augment: Learning Augmentation Policies from Data”, IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019 [last revised 11 Apr 2019 (v3)]

[0006] The automatic data augmentation technology disclosed in Non-Patent Document 1 uses discrimination accuracy to calculate the objective function, and since discrimination accuracy is repeatedly calculated while changing the data augmentation method, the calculation cost becomes excessive.

[0007] One objective of the present disclosure is to provide an information processing device and an information processing method that optimize a data set while suppressing calculation costs.

[0008] An information processing device according to one aspect of the present disclosure has a memory that stores a program and a processor that executes processing in accordance with the program, and by executing the program, the processor generates an updated dataset by adding or deleting data from a reference dataset in accordance with a strategy, converts the updated dataset into converted data that is data for inputting the updated dataset into a regression model that estimates the discrimination accuracy of the updated dataset using a discriminant model, and determines the strategy based on an evaluation value of an objective function that includes the discrimination accuracy estimated by the regression model.

[0009] According to one aspect of the present disclosure, when calculating the evaluation value of an objective function, the discrimination accuracy of a data set can be estimated using a regression model, thereby eliminating the need to learn a discrimination model with a data set each time, thereby reducing computational costs.

[0010] 1. A block diagram showing an example of the configuration of a dataset optimization device according to an embodiment. 2. A block diagram showing an example of the hardware configuration of the dataset optimization device shown in FIG. 1. 3. A diagram schematically showing the flow of data in the dataset optimization unit shown in FIG. 1. 4. A diagram for explaining a learning method of a regression model used in the estimation unit shown in FIG. 1. 5. A flowchart showing an example of the operation procedure of the dataset optimization device according to an embodiment. 6. A flowchart showing a specific example of the processing of step S104 shown in FIG. 5. 7. A diagram schematically showing the flow of data in the objective function calculation unit shown in FIG. 1. 8. A flowchart showing a specific example of the processing of step S201 shown in FIG. 9. 10. A diagram showing an example of a method for creating converted data when text data is not used in the determination of step S402 shown in FIG. 10. 11. A diagram showing an example of a method for creating converted data when text data is used in the determination of step S402 shown in FIG. 10. 12. A diagram showing an example of a case where data is sequentially added to a trained dataset. 13. A diagram showing an example of a case where some data is replaced or some data is deleted from a trained dataset. 14. A graph showing an example of a calculation cost displayed on an output device. 15. A diagram showing an example of converted data displayed on an output device.

[0011] The dataset optimization device of this embodiment efficiently creates a dataset that is optimal for the target machine learning by repeatedly updating a dataset that includes multiple pieces of data. An example of the dataset optimization device of this embodiment will be described below.

[0012] The configuration of a dataset optimization device according to this embodiment will be described with reference to the drawings. FIG. 1 is a block diagram showing an example configuration of a dataset optimization device according to this embodiment. The dataset optimization device 1 includes a storage device 2 and a dataset optimization unit 3. The dataset optimization unit 3 includes a policy learning unit 5, a data updating unit 6, and an objective function calculation unit 4. The objective function calculation unit 4 includes a data conversion unit 7, an estimation unit 8, a discriminant model learning unit 9, a determination unit 10, an objective function processing unit 11, and a display processing unit 12.

[0013] When a reference dataset, which is the original dataset, is input, the policy learning unit 5 stores the reference dataset in the storage device 2. The policy learning unit 5 determines a policy, which is an update policy for the reference dataset, and instructs the determined policy to the data updating unit 6. The policy learning unit 5 determines a policy so as to obtain an optimal dataset, which is a dataset that maximizes the evaluation value Fv of the objective function received from the objective function calculation unit 4. The policy learning unit 5 outputs the optimal dataset and an optimal discriminant model. The optimal discriminant model is a discriminant model when the optimal dataset is obtained.

[0014] The data updating unit 6 generates an updated data set by adding or deleting data to the reference data in accordance with the policy instructed by the policy learning unit 5. The data updating unit 6 transmits the updated data set to the data conversion unit 7 or the discriminant model learning unit 9 in accordance with the destination information instructed by the policy learning unit 5.

[0015] The determination unit 10 determines whether to use a discriminant model to find the discrimination accuracy for the updated data set or a regression model to estimate the discrimination accuracy, and determines which model to use to evaluate the discrimination accuracy. If the determination unit 10 determines that the regression model should be used to evaluate the updated data set, the data conversion unit 7 converts the updated data into data to be input to the regression model. Hereinafter, the data to be input to the regression model will be referred to as converted data.

[0016] When the transformed data is input, the estimation unit 8 applies the transformed data to a regression model to estimate the discrimination accuracy. The estimation unit 8 transmits the estimated discrimination accuracy to the objective function processing unit 11. When an updated data set is input, the discriminant model learning unit 9 applies the updated data set to the discriminant model to calculate the discrimination accuracy. The discriminant model is a learning model that performs class classification. The discriminant model learning unit 9 transmits the calculated discrimination accuracy to the objective function processing unit 11. When the objective function processing unit 11 acquires the discrimination accuracy estimated by the regression model or the discrimination accuracy calculated by the discriminant model, it calculates an objective function evaluation value Fv including the discrimination accuracy. The objective function processing unit 11 transmits the objective function evaluation value Fv to the policy learning unit 5. The display processing unit 12 displays the processing time and the transformed data when calculating the objective function evaluation value Fv on a display (not shown). The processing time is the time required for calculating the objective function evaluation value Fv. The calculation cost of the objective function is proportional to the processing time. The display processing unit 12 may display the calculation cost instead of the processing time on a display (not shown).

[0017] 2 is a block diagram showing an example of the hardware configuration of the dataset optimization device shown in FIG. 1. The dataset optimization device 1 is, for example, an information processing device such as a personal computer. The dataset optimization device 1 has a storage device 2, a memory 22 that stores a program, and a processor 21 that executes processing in accordance with the program. The dataset optimization device 1 also has an input device 23, an output device 24, a communication device 25, a reading device 27, a connection port 28, and a drive 29. The storage device 2, the processor 21, the memory 22, the input device 23, the output device 24, the communication device 25, the reading device 27, the connection port 28, and the drive 29 are connected to each other via a bus 26 so as to be able to communicate with each other.

[0018] The storage device 2 is a storage device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive). The storage device 2 stores a program group including a plurality of programs, a data group including a plurality of types of data, and a learning model group including a plurality of learning models. The program group includes a policy learning program, a data update program, a conversion data creation program, a regression model learning program, a reliability evaluation program, a discriminant model learning program, an objective function processing program, and a display processing program. The data group includes a reference data set, an updated data set, an optimal data set, conversion data, and processing time data. The learning model group includes an optimization discriminant model, a policy learning model, a regression model, and a discriminant model.

[0019] The reference dataset and the updated dataset each contain multiple pieces of data and information on the number of times L that each piece of data has been used for learning. The number of times L that each piece of data has been used for learning will now be explained. To simplify the explanation, consider a case where a dataset contains three pieces of data: data A, data B, and data C. Assume that there are two combinations of data used for learning the learning model: a combination of data A and data B, and a combination of data A to data C. In this case, the number of times L that each piece of data A and data B has been learned is 2, but the number of times L that data C has been learned is 1. Within a dataset, data with a large number of times L that each piece of data has been learned is highly important for machine learning.

[0020] The processor 21 is an arithmetic device such as a CPU (Central Processing Unit), an MPU (Micro Processing Unit), or a GPU (Graphics Processing Unit). The memory 22 is a non-volatile memory such as a flash memory. The memory 22 reads out from the storage device 2 and stores therein programs to be executed by the processor 21. Various programs stored in the storage device 2 are read from the storage device 2 into the memory 22, and the programs are executed by the processor 21, thereby performing each function of the dataset optimization unit 3 shown in FIG. 1 .

[0021] The input device 23 is an input device such as a keyboard, mouse, or touch panel. The output device 24 is an output device such as a display that displays information or a speaker that outputs sound. In this embodiment, the output device 24 is a display. The communication device 25 has a communication circuit that communicates with other devices via a wired or wireless connection. The communication device 25 may be communicatively connected to other devices via the Internet. In this case, the communication device 25 has a communication circuit that transmits and receives data to and from other devices according to a communication protocol such as IP (Internet Protocol). The reading device 27 is, for example, a code reader that reads information from a one-dimensional code or a two-dimensional code. The connection port 28 is a terminal for communicatively connecting an external device to the bus 26. The drive 29 is a device that reads and writes data from and to the recording medium 30. The recording medium 30 is, for example, an optical disc such as a CD (Compact Disc) or a DVD (Digital Versatile Disc).

[0022] Note that some or all of the functions of the dataset optimization unit 3 may be performed by a dedicated circuit such as an ASIC (Application Specific Integrated Circuit). While FIG. 2 illustrates a case in which the storage device 2, processor 21, memory 22, input device 23, output device 24, communication device 25, reading device 27, connection port 28, and drive 29 are connected via a bus 26, the connection method is not limited to the bus 26. Some or all of the storage device 2, processor 21, memory 22, input device 23, output device 24, communication device 25, reading device 27, connection port 28, and drive 29 may be connected via a network. These components may be distributed across multiple information processing devices rather than being located in a single information processing device. Multiple processors 21 may be distributed, and the functions of the dataset optimization unit 3 may be executed by the multiple processors 21 working together. The network may be a local area network (LAN) or the Internet, or a combination of these networks. The location where the program to be loaded into the memory 22 is stored is not limited to the storage device 2, but may also be a portable recording medium 30.

[0023] Next, data processing in the dataset optimization unit 3 shown in Fig. 1 will be described in detail with reference to Fig. 3. Fig. 3 is a diagram schematically showing the flow of data in the dataset optimization unit shown in Fig. 1. Since Fig. 3 shows repeated policy learning to obtain an optimal dataset, the display processing unit 12 shown in Fig. 1 is omitted from Fig. 3.

[0024] The policy learning model used by the policy learning unit 5 is a learning model that repeatedly updates the dataset, calculates the objective function evaluation value Fv, and updates the policy to find a policy that maximizes the objective function evaluation value Fv. A policy generally refers to an action to be taken for a certain state. In this embodiment, the policy includes, in optimizing the dataset, the content of the transformation process to be performed on the data, the number of data to be generated, and the number of times the data is learned. The content of the transformation process includes, for example, rotating, cropping, or flipping an image. The policy learning unit 5 determines that the dataset has been optimized when the objective function evaluation value Fv is equal to or greater than a predetermined threshold Fth. The policy learning unit 5 may also determine that the updated dataset has been optimized when the number of times HC of the policy has been learned is equal to or greater than a predetermined threshold Hth.

[0025] In the policy learning cycle, if the discriminant model learning unit 9 learns a discriminant model every time the data set is updated, there is a problem that policy learning takes a long time. A case in which discrimination accuracy is obtained using only a discriminant model in policy learning will be referred to as a comparative example below. If the evaluation value of the objective function in the comparative example is Fv0, the evaluation value Fv0 is expressed by the formula "Fv0 = discrimination accuracy × α - number of data × β". The number of data is included in the negative loss function of the objective function. From this calculation formula, it can be seen that the smaller the number of data, the larger the evaluation value Fv0.

[0026] On the other hand, in this embodiment, when calculating the discrimination accuracy in policy learning, a regression model is used instead of a discriminant model, thereby reducing calculation costs. A regression model is a learning model that uses a dataset trained with a discriminant model as a dependent variable and estimates the discrimination accuracy as a response variable. In this embodiment, a discriminant model or a regression model is used when obtaining the discrimination accuracy from an updated dataset. If the evaluation value of the objective function in this embodiment is Fv, the evaluation value Fv is expressed by the formula "Fv = discrimination accuracy × α - number of data × β - total number of learning times × γ". The total number of learning times is the sum of the number of learning times L for each data. In other words, the total number of learning times is the sum of the number of times each data is input to the discriminant model when learning with the discriminant model. From this calculation formula, it can be seen that the smaller the number of data and / or the total number of learning times, the larger the evaluation value Fv. The objective function processing unit 11 calculates the evaluation value Fv using this calculation formula. Dataset optimization is the creation of a dataset that simultaneously maximizes the discrimination accuracy of the discriminant model and minimizes the learning cost (number of data and total number of learning times).

[0027] The computational cost of policy learning in the comparative example is the "learning cost of the discriminant model." In contrast, the computational cost of policy learning in this embodiment is the "learning cost of the discriminant model + learning cost of the regression model." However, in this embodiment, part of the discrimination accuracy is estimated by the regression model, so the learning cost of the discriminant model is significantly reduced. Because the learning cost of the regression model is small, the total computational cost of this embodiment is reduced compared to the comparative example.

[0028] As described above, the regression model used by the estimation unit 8 is a learning model that estimates the discrimination accuracy as a response variable using a dataset learned using a discriminant model as a dependent variable. The determination unit 10 determines whether the reliability of the regression model is high. If the determination unit 10 determines that the reliability of the regression model is low, the discriminant model learning unit 9 calculates the discrimination accuracy using the discriminant model. For example, if the dataset includes data that has not been learned in the discriminant model, the reliability of the regression model is determined to be low. If the reliability of the regression model is determined to be low, the estimation unit 8 uses the discrimination accuracy calculated using the discriminant model as training data for the regression model. On the other hand, if the determination unit 10 determines that the reliability of the regression model is high, the estimation unit 8 estimates the discrimination accuracy using the regression model. As such, this embodiment is a hybrid method that combines a discriminant model and a regression model. In the hybrid method, if the reliability of the regression model is high, the discrimination accuracy is estimated using the regression model, thereby reducing calculation costs compared to calculating the discrimination accuracy using only the discriminant model.

[0029] The criterion for determining the reliability of the regression model is, for example, an index value Uv of the uncertainty of estimation by the regression model or the number of times KC of learning of the discriminant model. The determination unit 10 determines that the reliability of the regression model is high when the index value Uv of the uncertainty is equal to or less than a predetermined threshold Uth, and determines that the reliability of the regression model is low when the index value Uv of the uncertainty is greater than the threshold Uth. Furthermore, the determination unit 10 determines that the reliability of the regression model is high when the number of times KC of learning of the discriminant model is equal to or greater than a predetermined threshold Kth, and determines that the reliability of the regression model is low when the number of times KC of learning is less than the threshold Kth.

[0030] A specific example of a method for evaluating the reliability of estimations made using a regression model will be described. Evaluation methods include, for example, (a) Monte Carlo dropout, (b) Bayesian neural networks, and (c) ensemble models. (a) Monte Carlo dropout is a method for evaluating the magnitude of fluctuations in inference by dropping out a model multiple times. (b) Bayesian neural networks are a method for calculating an index value for prediction uncertainty using a neural network in which the weights themselves are random variables. (c) Ensemble models are a method for training multiple models and calculating an index value for uncertainty from the distribution of predictions from each model.

[0031] In this embodiment, when obtaining the evaluation value Fv of the objective function of the updated data set, the discrimination accuracy can be estimated using a regression model, so there is no need to learn a discrimination model every time the data set is updated, and calculation costs can be reduced.

[0032] In this embodiment, the determining factors of the policy in the policy learning unit 5 are not limited to the content of the transformation of the data set. For example, the determining factors of the policy may include the number of times KC the discriminant model has been learned. In this case, in the policy, the number of times KC the discriminant model has been learned is reflected in the reliability of the discrimination accuracy by the regression model.

[0033] Next, a method for learning a regression model will be described with reference to Fig. 4. Fig. 4 is a diagram for explaining the method for learning a regression model used in the estimation unit shown in Fig. 1. The estimation unit 8 learns the regression model using, as training data, a combination of a trained data set and discrimination accuracy obtained in the process of learning the discriminant model each time the data set is updated.

[0034] Each of multiple data sets, data set Ds1 to data set Dsn (n is an integer greater than or equal to 2), is used as an input to the discriminant model, and discrimination accuracy da1 to discrimination accuracy dan are used as the output of the discriminant model. Transformed data Dt1 to Dtn are obtained for each of data sets Ds1 to Dsn. Data set Ds learned using the discriminant model is converted, for example, into fixed-size two-dimensional image data, and this data is used as transformed data Dt to be input to the regression model. If the combination of input and output of the learning model is expressed as (input, output), a regression model is obtained using each combination of (transformed data Dt1, discrimination accuracy da1) to (transformed data Dtn, discrimination accuracy dan) as training data. The estimation unit 8 can train the regression model using a convolutional neural network (CNN) and estimate the discrimination accuracy. While the description has been given for a case where the number of training data is two or more, the number of training data may be one or more.

[0035] The estimation unit 8 can estimate the discrimination accuracy using a trained regression model, even for a dataset in which unknown data has been added to a trained dataset, or a dataset in which some data has been deleted from a trained dataset. Referring to FIG. 4 , the combination of the explanatory variables and objective variables of the discriminant model serves as training data for the regression model, so the greater the number of times KC the discriminant model has been trained, the higher the reliability of the regression model. When a trained regression model can be used, the discrimination accuracy can be estimated using the regression model without having to train the updated dataset with the discriminant model from scratch, allowing for efficient policy learning. When calculating discrimination accuracy as an objective function in the policy learning cycle, learning a discriminant model is not necessary, thereby reducing the computational costs of learning.

[0036] Next, the operation of the dataset optimization device 1 of this embodiment will be described. FIG. 5 is a flowchart showing an example of the operation procedure of the dataset optimization device according to the embodiment. In step S101, when a reference dataset is input via the input device 23, the policy learning unit 5 stores the reference dataset in the storage device 2. In step S102, the policy learning unit 5 transmits a policy including information on the data conversion method and the number of learning times to the data updating unit 6. If this is the first time, the policy learning unit 5 sets a random value as the initial value of the number of learning times. In step S103, the data updating unit 6 updates the reference dataset based on the policy and creates an updated dataset. In step S104, the objective function calculation unit 4 calculates an objective function evaluation value Fv from the updated dataset.

[0037] In step S105, the policy learning unit 5 updates the policy based on the evaluation value Fv of the objective function calculated in step S104. In step S106, the policy learning unit 5 determines whether the condition that the evaluation value Fv of the objective function is equal to or greater than a threshold Fth or the condition that the number of times HC of policy learning is equal to or greater than a threshold Hth is satisfied. If the result of the determination in step S106 is that neither condition is satisfied, the policy learning unit 5 returns to the processing in step S102. On the other hand, if the result of the determination in step S106 is that at least one of the conditions is satisfied, the policy learning unit 5 uses the final updated dataset as an optimal dataset and learns the discriminant model (step S107). In step S108, the policy learning unit 5 outputs the optimal dataset and the optimal discriminant model via the output device 24.

[0038] Next, the operation of the objective function calculation unit 4 in step S104 shown in Fig. 5 will be described in detail. Fig. 6 is a flowchart showing a specific example of the process of step S104 shown in Fig. 5. Fig. 7 is a diagram schematically showing the flow of data in the objective function calculation unit shown in Fig. 1. In Fig. 7, the same step numbers are assigned to processes that are the same as those shown in Fig. 6.

[0039] In step S201 shown in FIG. 6 , the data conversion unit 7 creates converted data when the updated data set created by the data update unit 6 and the data of the number of learning times KC are input. The specific processing of step S201 will be described later. In step S202, the determination unit 10 determines whether the updated data set is the first. If it is the first updated data set, the determination unit 10 initializes the regression model and the discriminant model (step S203). In step S204, the discriminant model training unit 9 trains a discriminant model using the updated data set as training data and calculates the discrimination accuracy. The updated data set is input to the discriminant model training unit 9 as training data for the discriminant model (step S301 shown in FIG. 7 ). In step S205, the estimation unit 8 trains a regression model using the converted data as a dependent variable and the discrimination accuracy calculated by the discriminant model training unit 9 as a response variable. The discrimination accuracy is input to the estimation unit 8 as training data for the regression model (step S302 shown in FIG. 7 ).

[0040] In step S206, the objective function processing unit 11 calculates an evaluation value Fv of the objective function based on the discrimination accuracy, the number of data, and the total number of learning times. The objective function processing unit 11 transmits the calculated evaluation value Fv to the policy learning unit 5. In step S207, the display processing unit 12 displays the calculation cost and conversion data for calculating the evaluation value Fv of the objective function on the output device 24.

[0041] On the other hand, if the result of the determination in step S202 indicates that the updated data set is not the first, the determination unit 10 calculates an index value Uv of the uncertainty of the estimation by the regression model (step S208). For example, when evaluating the reliability of the estimation of the discrimination accuracy by the regression model using the Monte Carlo dropout method, the determination unit 10 inputs the converted data into the regression model, performs dropout in the regression model, and calculates the uncertainty index value Uv from the inference variability. In step S209, the determination unit 10 determines whether the index value Uv of the uncertainty of the estimation by the regression model is equal to or less than a threshold Uth, or whether the number of learning cycles KC of the discrimination model is equal to or greater than a threshold Kth. If the result of the determination in step S209 indicates that neither condition is met, the objective function calculation unit 4 proceeds to the processing of step S204.

[0042] On the other hand, if the result of the determination in step S209 indicates that either of the conditions is satisfied, the estimation unit 8 inputs the converted data into the regression model and estimates the discrimination accuracy (step S210). If the index value Uv of the uncertainty of the estimation by the regression model is equal to or less than the threshold Uth, the reliability of the estimation by the regression model is considered to be high. Furthermore, if the number of times KC the discriminant model has been learned is equal to or greater than the threshold Kth, the reliability of the estimation by the regression model trained using the discrimination results of the discriminant model is considered to be high. Therefore, by estimating the discrimination accuracy using the regression model, calculation costs are reduced. After step S210, the objective function calculation unit 4 proceeds to the processing of step S206.

[0043] Next, the operation of the data conversion unit 7 in step S201 shown in Fig. 6 will be described in detail. Fig. 8 is a flowchart showing a specific example of the process of step S201 shown in Fig. 6. Fig. 9 is a diagram showing an example of a method for creating converted data when text data is not used in the determination of step S402 shown in Fig. 8. Fig. 10 is a diagram showing an example of a method for creating converted data when text data is used in the determination of step S402 shown in Fig. 8.

[0044] In step S401, the updated data set is input to the data conversion unit 7. In step S402, the data conversion unit 7 determines whether or not to use text data for each piece of data in the updated data set. If each piece of data in the updated data set is two-dimensional image data, the text data is image caption data. If a car driving on a white road is displayed in the image, "car driving on a white road" is the text data. If the data conversion unit 7 does not use text data, it proceeds to processing in step S403; if it uses text data, it proceeds to processing in step S404.

[0045] In step S403, the data conversion unit 7 performs dimensional compression on each piece of data. After step S403, the data conversion unit 7 proceeds to the processing of step S406. In step S406, the data conversion unit 7 converts the compressed data into 2D×K pieces of image data. The data conversion unit 7 converts the number of times Lj (j is an arbitrary integer from 1 to n) of learning for each piece of data included in the updated data set into a luminance value. When the luminance value is the minimum value, the displayed color is black, and when the luminance value is the maximum value, the displayed color is white. The luminance values ​​are ranked into, for example, 255 gradations between the maximum and minimum values. In step S407, the data conversion unit 7 combines the K pieces of two-dimensional image data to convert the K pieces of two-dimensional image data into one piece of two-dimensional image data.

[0046] On the other hand, if the determination result of step S402 in FIG. 8 indicates that text data is to be used, the data conversion unit 7 generates corresponding text data from each data item in the updated data set (step S404). In step S405, the data conversion unit 7 vectorizes the generated text data. For example, the data conversion unit 7 converts the text data into vector data representing attribute information of the text data. Attribute information is information about attributes, which are properties that characterize data. Attributes include, for example, the type, shape, or size of an object. The attribute information includes attribute values, which are values ​​indicating the degree of the attribute. If the attribute is the color of an object, the attribute value is, for example, red. If the attribute is the shape of an object, the attribute value is, for example, cube. After step S405, the data conversion unit 7 proceeds to processing of step S406. In step S406, the data conversion unit 7 converts the vector data into 2D×K image data. The data conversion unit 7 converts the number of learning times Lj of each data item included in the updated data set into a brightness value. After step S406, the data conversion unit 7 proceeds to processing of step S407.

[0047] The processing of steps S403, S406, and S407 will be described with reference to FIG. 9. FIG. 9 is a diagram showing converted data when using latent variables of an unsupervised learning model. A dataset has n pieces of data consisting of N-dimensional data 1 to data n. N in N dimensions is, for example, 2. When N=2, data 1 to data n are two-dimensional image data. The number of times data 1 to data n are learned is L1 to Ln, respectively. The data conversion unit 7 dimensionally compresses the N-dimensional data 1 to data n into M-dimensional vector data 1 to vector data n. The vector data is a latent variable that indicates features extracted from the data.

[0048] The data conversion unit 7 decomposes the M-dimensional vector data 1 to vector data n into (2 x k) dimensions and converts them into K pieces of image data. The data conversion unit 7 expresses the number of times Lj each piece of data has been learned as a brightness value. The data conversion unit 7 combines the K pieces of image data into one piece of image data. If the feature is two-dimensional, it becomes image data. By the data conversion unit 7 visualizing the converted data as two-dimensional data, the information in the dataset becomes a fixed data size, independent of the number of pieces of data in the dataset.

[0049] The data conversion unit 7 uses, for example, an autoencoder or statistical technology for dimensionality reduction. Examples of autoencoders include a variational autoencoder (VAE) or a multi-facet clustering VAE (MFCVAE). A VAE is an unsupervised learning model that learns the latent distribution of input data and has the ability to generate new data. An MFCVAE is a type of variational autoencoder that extends a variational autoencoder to enable output of latent variables from multiple perspectives. Examples of statistical technologies include t-distributed stochastic neighbor embedding (t-SNE). t-SNE is a technology that reduces high-dimensional data to two or three dimensions.

[0050] In the case of the converted data described with reference to Fig. 9, two-dimensional image data is generated using features acquired by unsupervised learning or statistical techniques. Because the converted data is created using unsupervised learning or statistical techniques for dimensionality compression, supervised data such as caption data is not required for each data.

[0051] The processing of steps S404 to S407 will be described with reference to Fig. 10. Fig. 10 is a diagram showing text data and conversion data when vector representation of text data is used.

[0052] The data conversion unit 7 generates text data from each of the N-dimensional data 1 to data n. FIG. 10 shows a case where the text data of data 1 to data n are caption data 1 to caption data n, respectively. The data conversion unit 7 vectorizes the N-dimensional caption data 1 to caption data n and converts them into M-dimensional vector data 1 to vector data n. The data conversion unit 7 then decomposes the M-dimensional vector data 1 to vector data n into (2×k) dimensions and converts them into K pieces of image data. The data conversion unit 7 represents the number of times Lj that each piece of data has been learned as a luminance value. The data conversion unit 7 combines the K pieces of image data into one piece of image data.

[0053] In the case of the converted data described with reference to FIG. 10, each vector data represents attribute information of the text, so that semantic information of the data is reflected in the estimation of discrimination accuracy, thereby improving the estimation accuracy.

[0054] 9 and 10 , circles in the two-dimensional images represent data with a small number of learning times Lj, and squares represent data with a large number of learning times Lj. Triangles in the two-dimensional images represent data with a learning time Lj between the circle and square. Since the learning time Lj of each data item is expressed by a luminance value, the determination unit 10 can grasp the degree of influence of each data item on the discrimination model from differences in luminance values. For example, data with a large number of learning times L is highly important for machine learning and has a large influence on discrimination accuracy.

[0055] While the case where the number of learning times of each data item in the converted data is expressed as a luminance value has been described with reference to FIGS. 9 and 10 , the uncertainty index value of the regression model may also be expressed as a luminance value. For example, an image indicating the uncertainty index value as a luminance value may be added to one of the K images. In this case, the determination unit 10 can determine the reliability of the regression model by referring to the luminance value of the regression model attached to the image data of the converted data each time a new updated data set is generated. Furthermore, in this embodiment, the case where the data conversion unit 7 converts the updated data set into K two-dimensional data items has been described, but the data is not limited to two dimensions and may be three-dimensional or four-dimensional.

[0056] Next, the advantages of this embodiment in calculating the discrimination accuracy, which is performed each time the data set is updated, will be described with reference to Fig. 11 and Fig. 12. Fig. 11 is a diagram showing an example of sequentially adding data to a trained data set. Fig. 12 is a diagram showing an example of replacing or deleting some data from a trained data set.

[0057] The updated dataset Dx1 shown in FIG. 11 is a dataset in which additional data 1 is added to the trained dataset. When additional learning of a small amount of data is performed, learning is performed each time data is added, and the discrimination accuracy is determined, so the required computational cost is low. Therefore, when training with the updated dataset Dx1, the training model trained with the trained dataset can be used. In this embodiment, the discriminant model training unit 9 simply inputs additional data 1 into the trained discriminant model and performs training.

[0058] 11 is a data set in which additional data 2 is added to the updated data set Dx1. Therefore, when learning with the updated data set Dx2, the learning model learned with the updated data set Dx1 can be used. Therefore, the discriminant model learning unit 9 can learn by inputting the additional data 2 to the already-learned discriminant model.

[0059] The updated dataset Dx3 shown in Figure 12 is a dataset in which some data has been deleted from the trained dataset. After a learning model has been trained using a certain dataset, it is not possible to return to a state in which some data has been deleted from the learning model. Therefore, it is necessary to restart training of the learning model using the updated dataset Dx3 after some data has been deleted from the trained data.

[0060] The updated dataset Dx4 shown in Figure 12 is a dataset in which some of the data in the trained dataset has been replaced with untrained data. The updated dataset Dx4 is a dataset in which additional data 3 has been added in place of the deleted data in the updated dataset Dx3. The updated dataset Dx4 is also an untrained dataset, just like the updated dataset Dx3. Therefore, it is necessary to redo the training of the training model using the updated dataset Dx4 in which some of the trained data has been replaced.

[0061] In this embodiment, in response to the problem described with reference to Fig. 12, the discrimination accuracy is estimated using a regression model for the updated data sets Dx3 and Dx4. This eliminates the need to retrain the discrimination model, thereby reducing calculation costs. This will be described with reference to the calculation formula for the evaluation value Fv (Fv = discrimination accuracy × α - number of data × β - total number of learning times × γ).

[0062] For example, if an evaluation value Fv is calculated for the updated data set Dx3 using only the discriminant model, the number of data in the above calculation formula will be equivalent to the number of data in the trained data set. On the other hand, if the evaluation value Fv is calculated using the discrimination accuracy estimated by the regression model, the number of data in the above calculation formula will be significantly smaller than the number of data in the updated data set Dx3. Therefore, it can be seen that the evaluation value Fv in this embodiment is large and the calculation cost is reduced.

[0063] Next, the data displayed on the output device 24 in step S207 shown in Fig. 6 will be described. Fig. 13 is a graph showing an example of the calculation cost displayed on the output device. The horizontal axis of Fig. 13 represents the number of calculations of the evaluation value Fv of the objective function, and the vertical axis represents the calculation cost. In Fig. 13, the present embodiment, in which a discriminant model and a regression model are combined, is shown by a solid bar graph, and a comparative example in which only a discriminant model is used is shown by a dashed bar graph. The vertical axis of Fig. 13 may be replaced with processing time.

[0064] When the number of updates to the dataset is small, such as the first time the evaluation value Fv is calculated, the reliability of the regression model estimation is low in this embodiment, and a discriminant model is used. Therefore, the processing time and calculation cost are equivalent in this embodiment and the comparative example. When the number of data in the dataset increases in the second time the evaluation value Fv is calculated, the processing time and calculation cost are both shorter in this embodiment and the comparative example. This is as explained with reference to FIG. 11 .

[0065] On the other hand, when the number of data in the dataset decreases, such as after the third calculation of the evaluation value Fv, as described with reference to FIG. 12 , in the comparative example, the discriminant model is used and learning is restarted from the beginning using an updated dataset. In contrast, in this embodiment, a regression model is used, so the processing time of this embodiment is shorter than that of the comparative example. As a result, the calculation cost of this embodiment is lower than that of the comparative example. By displaying the calculation cost in this manner, the user can visually recognize that the calculation cost of this embodiment is lower than that of the comparative example. Note that while FIG. 13 shows the case where the calculation cost of the comparative example is displayed, the calculation cost of the comparative example does not necessarily have to be displayed.

[0066] Fig. 14 is a diagram showing an example of transformed data displayed on an output device. The horizontal axis of Fig. 14 represents the number of calculations of the evaluation value Fv of the objective function, and the vertical axis represents an image of two-dimensional image data of the transformed data. Circles, triangles, and squares displayed on each image represent the number of learning iterations, as in Figs. 9 and 10. Although not shown in Fig. 14, the circles, triangles, and squares are represented by different brightness values ​​(colors).

[0067] As shown in FIG. 14 , as the number of calculations of the evaluation value Fv increases, the number of data with a high number of learning times decreases. In this embodiment, the evaluation value Fv is calculated using a small number of data and a small total number of learning times. By displaying the converted data, the user can visually recognize the converted data. Furthermore, if the number of learning times is expressed as a brightness value in the displayed converted data, the user can visually recognize the learning frequency of each data.

[0068] Next, the effects of this embodiment will be explained by comparing the calculation costs of a comparative example with those of this embodiment. The comparative example is a case where only a discriminant model is used to obtain discrimination accuracy in a policy learning cycle. This embodiment is a case where a discriminant model and a regression model are combined as models used to obtain discrimination accuracy in a policy learning cycle. First, various calculation costs Ca(x) to Cc(x) and C1(x) to C9(x) related to basic calculations are defined as follows:

[0069] Ca(x) is the calculation cost when calculating the evaluation value of the objective function x times in the comparative example, where x is any integer equal to or greater than 1. Cb(x) is the calculation cost when using latent variables of an unsupervised learning model as converted data and calculating the evaluation value of the objective function x times in the present embodiment. Cc(x) is the calculation cost when using text data as converted data and calculating the evaluation value of the objective function x times in the present embodiment.

[0070] C1(x) is the computational cost when learning a discriminant model x times. C2(x) is the computational cost when learning an unsupervised learning model for extracting features x times. Even if the evaluation value of the objective function is calculated multiple times, the unsupervised learning model only needs to be learned once. C3(x) is the computational cost when learning a regression model x times. C4(x) is the computational cost when converting data used for learning into vector data (latent variables) x times. C5(x) is the computational cost when converting from vector data to image data x times. C6(x) is the computational cost when estimating discrimination accuracy from image data x times. C7(x) is the computational cost when evaluating the reliability of estimation using a regression model x times. C8(x) is the computational cost when creating text data x times from data used for learning. C9(x) is the calculation cost when creating vector data from text data x times.

[0071] Under the above definition, in the comparative example, the calculation cost when calculating the evaluation value of the objective function N times is expressed by the following formula (1): Ca(N)=C1(N) (1)

[0072] In this embodiment, the calculation cost when using the latent variables of the unsupervised learning model for the converted data and calculating the evaluation value of the objective function N times is expressed by formula (2). However, it is assumed that M times of estimation are performed using the regression model, and M is sufficiently large (M>>0). Cc(N) = C1(N-M) + C2(1) + C3(M) + C4(N) + C5(N) + C6(N) + C7(N) ... (2)

[0073] In this embodiment, the calculation cost when using text data as the converted data and calculating the evaluation value of the objective function N times is expressed by equation (3). However, it is assumed that estimation is performed M times using a regression model, and M is sufficiently large (M>>0). Cb(N)=C1(N-M)+C3(M)+C5(N)+C6(N)+C7(N)+C8(N)+C9(N) ... (3)

[0074] Generally, the relationship is C1(x), C2(x)>C3(x)>>C4(x), C5(x), C6(x), C7(x), C8(x), C9(x). From this relationship and equations (1) to (3), equations (4) and (5) are derived. Ca(N)>Cb(N) ... (4) Ca(N)>Cc(N) ... (5)

[0075] From equations (4) and (5), it can be seen that the calculation cost of this embodiment is smaller than that of the comparative example. When M is close to 0, the calculation cost of the comparative example is small, but M is generally very large. The larger the value of M, the smaller the calculation cost.

[0076] In policy learning, data augmentation or data reduction is required to prepare a dataset that maximizes the discrimination accuracy of a discriminant model for class classification with as little training data as possible. The automatic data augmentation technology disclosed in Non-Patent Document 1 uses discrimination accuracy to calculate the objective function, and because discrimination accuracy is repeatedly calculated while changing the data augmentation policy, computational costs become excessive. Furthermore, with the technology disclosed in Non-Patent Document 1, if learning is performed with all data combinations and discrimination accuracy is saved, the model storage capacity becomes enormous. With the technology disclosed in Non-Patent Document 1, even if a model can be saved with the patterns of all combinations of datasets, it is not possible to estimate the discrimination accuracy of unknown datasets. In contrast, in this embodiment, the discriminant model does not need to be learned with all updated datasets, thereby reducing computational costs.

[0077] The dataset optimization device 1 of this embodiment includes a memory 22 that stores a program, and a processor 21 that executes processing in accordance with the program. By executing the program, the processor 21 generates an updated dataset by adding or deleting data from a reference dataset in accordance with a measure, converts the updated dataset into conversion data that is data for inputting the updated dataset into a regression model that estimates the discrimination accuracy of the updated dataset using a discriminant model, and determines a measure based on an evaluation value Fv of an objective function that includes the discrimination accuracy estimated by the regression model.

[0078] According to this embodiment, when calculating the evaluation value of the objective function, the discrimination accuracy of the data set can be estimated using a regression model, so there is no need to learn a discriminant model with the data set every time, which reduces calculation costs. As a result, the workload of the data preprocessing step in the AI ​​development process can be significantly reduced compared to the comparative example.

[0079] The operation of the dataset optimization device 1 of this embodiment will be specifically described. The discriminant model is a learning model that calculates the discrimination accuracy of class classification for each dataset. The discriminant model assumes that the dataset will be used for machine learning, and outputs the discrimination accuracy of the target dataset when each data of the dataset is input. On the other hand, the regression model is a model that estimates the discrimination accuracy of a dataset using converted data obtained by collectively converting the dataset, without inputting each data of the dataset. The discrimination accuracy estimated by the regression model is less accurate than that of the discriminant model, but has the advantage that the computational cost of the regression model is lower than that of the discriminant model. When calculating the discrimination accuracy of a newly updated dataset, either a discriminant model or a regression model is used. When the reliability of the regression model is high, the regression model is adopted. On the other hand, when the reliability of the regression model is low, the discriminant model is adopted. In the policy learning of this embodiment, the evaluation value Fv of the objective function is calculated each time the dataset is updated. If the discrimination accuracy is estimated using a regression model for some updated datasets among the repeatedly updated datasets, the computational cost of the evaluation value Fv can be reduced.

[0080] In this embodiment, the determination unit 10 determines whether to use a regression model or a discriminant model for the updated data set based on the reliability of the regression model. If the reliability of the regression model is high, the regression model is used to estimate the discrimination accuracy, and if the reliability of the regression model is low, the discriminant model is used to determine the discrimination accuracy. This reduces the calculation cost while maintaining the discrimination accuracy of the class classification.

[0081] The above-described embodiments are merely illustrative examples of the present invention, and are not intended to limit the scope of the present invention to these embodiments alone. Those skilled in the art can implement the present invention in various other forms without departing from the scope of the present invention.

[0082] The above-described embodiment also includes the following features, but the features included in the present embodiment are not limited to those listed below.

[0083] (Item 1) An information processing device comprising: a memory for storing a program; and a processor for executing processing in accordance with the program, wherein the processor, by executing the program, generates an updated dataset by adding or deleting data from a reference dataset in accordance with a measure; converts the updated dataset into conversion data that is data for inputting the updated dataset into a regression model that estimates the discrimination accuracy of the updated dataset using a discriminant model; and determines the measure based on an evaluation value of an objective function that includes the discrimination accuracy estimated by the regression model.

[0084] (Item 2) The information processing device according to item 1, wherein the processor includes, as a determining factor of the strategy, the number of times the discriminant model has been trained using the updated data set.

[0085] (Item 3) The information processing device according to item 1 or 2, wherein the processor performs dimensional compression on the updated data set to obtain a plurality of two-dimensional image data as the transformation data.

[0086] (Item 4) In the information processing device according to any one of items 1 to 3, the processor obtains, as the conversion data, a plurality of two-dimensional image data by vectorizing text data of each data of the updated data set.

[0087] (Item 5) In the information processing device according to any one of items 1 to 4, the processor represents, in the image data of the converted data, an index value of the number of times each data item in the updated data set has been learned or an index value of the uncertainty of the estimation by the regression model, as a brightness value.

[0088] (Item 6) In the information processing device according to any one of items 1 to 5, the processor determines whether to use the regression model or the discriminant model for the updated data set based on the reliability of the regression model.

[0089] (Item 7) In the information processing device according to item 6, when an index value of uncertainty in estimation by the regression model is equal to or less than a predetermined threshold, the information processing device estimates the discrimination accuracy of the updated data set using the regression model, and when the index value is greater than the threshold, the information processing device calculates the discrimination accuracy of the updated data set using the discriminant model.

[0090] (Item 8) In the information processing device described in item 6, the processor estimates the discrimination accuracy of the updated data set using the regression model when the number of times the discriminant model has been trained is equal to or greater than a predetermined threshold, and calculates the discrimination accuracy of the updated data set using the discriminant model when the number of times the discriminant model has been trained is less than the threshold.

[0091] (Item 9) In the information processing device described in item 6, the processor increases the evaluation value as the discrimination accuracy of the updated data set increases, and increases the evaluation value as the number of data in the updated data set and / or the total number of times the discrimination model has been trained decreases.

[0092] (Item 10) The information processing device according to any one of items 1 to 9, further comprising a display that displays information, and the processor causes the display to display the calculation cost when calculating the evaluation value and the conversion data.

[0093] 1 Dataset optimization device, 2 Storage device, 3 Dataset optimization unit, 4 Objective function calculation unit, 5 Policy learning unit, 6 Data update unit, 7 Data conversion unit, 8 Estimation unit, 9 Discriminant model learning unit, 10 Determination unit, 11 Objective function processing unit, 12 Display processing unit, 21 Processor, 22 Memory, 23 Input device, 24 Output device, 25 Communication device, 26 Bus, 27 Reading device, 28 Connection port, 29 Drive, 30 Recording medium.

Claims

1. An information processing device having: a memory for storing a program; and a processor for executing processing in accordance with the program, wherein the processor, by executing the program, generates an updated dataset by adding or deleting data from a reference dataset in accordance with a strategy; converts the updated dataset into converted data that is data for inputting the updated dataset into a regression model that estimates the discrimination accuracy of the updated dataset using a discriminant model; and determines the strategy based on an evaluation value of an objective function that includes the discrimination accuracy estimated by the regression model.

2. An information processing device according to claim 1, wherein the processor includes the number of times the discriminant model has been trained using the updated data set as a determining factor for the strategy.

3. An information processing device according to claim 1, wherein the processor performs dimensional compression on the updated data set to obtain a plurality of two-dimensional image data as the transformation data.

4. An information processing device according to claim 1, wherein the processor obtains a plurality of two-dimensional image data as the conversion data by vectorizing text data of each data item in the updated data set.

5. An information processing device according to claim 1, wherein the processor expresses the number of times each data item in the updated data set has been learned or an index value of the uncertainty of the estimation by the regression model in the image data of the converted data as a brightness value.

6. An information processing device according to claim 1, wherein the processor determines whether to use the regression model or the discriminant model for the updated data set based on the reliability of the regression model.

7. An information processing device according to claim 6, wherein the processor estimates the discrimination accuracy of the updated data set using the regression model when an index value of the uncertainty of the estimation by the regression model is equal to or less than a predetermined threshold, and calculates the discrimination accuracy of the updated data set using the discrimination model when the index value is greater than the threshold.

8. An information processing device according to claim 6, wherein the processor estimates the discrimination accuracy of the updated data set using the regression model when the number of times the discriminant model has been trained is equal to or greater than a predetermined threshold, and calculates the discrimination accuracy of the updated data set using the discriminant model when the number of times the discriminant model has been trained is less than the threshold.

9. An information processing device according to claim 6, wherein the processor increases the evaluation value as the discrimination accuracy of the updated data set increases, and increases the evaluation value as the number of data in the updated data set and / or the total number of times the discrimination model has been trained decreases.

10. An information processing device according to claim 1, further comprising a display for displaying information, wherein the processor causes the display to display the calculation cost for determining the evaluation value and the conversion data.

11. An information processing method executed by an information processing device, comprising: generating an updated dataset by adding or deleting data from a reference dataset according to a strategy; converting the updated dataset into converted data that is data for inputting the updated dataset into a regression model that estimates the discrimination accuracy of the updated dataset using a discriminant model; and determining the strategy based on an evaluation value of an objective function that includes the discrimination accuracy estimated by the regression model.

Images