Learning device, learning method, and learning program

Abstract

This training device calculates a degree of deviation La between the distribution of a first feature vector, which is a feature vector of training environment data in an intermediate layer of a deep learning model, and the distribution of a second feature vector, which is a feature vector of operation environment data in the intermediate layer of the deep learning model. Further, the training device calculates a loss LIM relating to the reliability of a value output by the deep learning model receiving input of the operation environment data. Then, the training device updates a parameter θ of a feature extractor fθ that extracts a feature vector in the deep learning model so as to reduce the sum of the degree of deviation La and the loss LIM relating to the reliability of the value output by the deep learning model as much as possible.

Description

[Technical field]

[0001] The present invention relates to a learning device, a learning method, and a learning program for a deep learning model. [Background technology]

[0002] In machine learning, if the distribution of input data when the model is in operation differs from the distribution of input data when the model is trained, the accuracy of the model decreases.

[0003] To solve such problems, a method has been proposed in which a new supervised dataset is created in the model's operating environment and additional model training is performed (Prior Art 1).In addition, a method has been proposed in which a model is trained using data from a supervised learning environment and data from an unsupervised operating environment simultaneously, creating a model that is less susceptible to the effects of differences in data from different environments (Prior Art 2). [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Bridle, John S., et al. "Unsupervised Classifiers, Mutual Information and 'Phantom Targets'", Advances in Neural Information Processing Systems 4 (NIPS 1991) [Non-Patent Document 2] Hendrycks, Dan, et al. "BENCHMARKING NEURAL NETWORK ROBUSTNESS TO COMMON CORRUPTIONS AND PERTURBATIONS", Proceedings of the International Conference on Learning Representations (ICLR), 2019 Summary of the Invention [Problem to be solved by the invention]

[0005] However, Conventional Technology 1 requires the creation of a new supervised dataset in the model's operating environment, which results in high costs. In addition, Conventional Technology 2 may have difficulty in obtaining data from the operating environment during the model learning stage. Furthermore, using data from a different environment for model learning may be difficult from the perspective of security and privacy.

[0006] Therefore, an object of the present invention is to solve the above-mentioned problems and to prevent the accuracy of the model from decreasing even when the distribution of input data when the model is operated differs from the distribution of input data when the model is trained. [Means for solving the problem]

[0007] In order to solve the above-mentioned problems, the present invention is characterized in comprising a first calculation unit that calculates a degree of mismatch between a distribution of a first feature vector, which is a feature vector in an intermediate layer of a deep learning model of training data used to train the deep learning model, and a distribution of a second feature vector, which is a feature vector in an intermediate layer of the deep learning model of data to be processed; a second calculation unit that calculates a loss related to the confidence of a value output by the deep learning model for the data to be processed; and an update unit that updates parameters of a feature extractor that extracts feature vectors in the deep learning model so that the sum of the value of the degree of mismatch and the loss related to the confidence is as small as possible. Effect of the Invention

[0008] According to the present invention, even if the distribution of input data when the model is operated differs from the distribution of input data when the model is trained, it is possible to suppress a decrease in the accuracy of the model. [Brief description of the drawings]

[0009] [Figure 1]FIG. 1 is a flowchart showing an outline of a processing procedure of the learning device. [Diagram 2] FIG. 2 is a diagram for explaining an outline of the learning device. [Diagram 3] FIG. 3 is a diagram illustrating an example of the configuration of a learning device. [Figure 4] FIG. 4 is a flowchart showing an example of a process (see S3 in FIG. 1) in which the learning device adapts a trained deep learning model to an operational environment. [Diagram 5] FIG. 5 is a flowchart showing an example of the processing procedure of S32 in FIG. [Figure 6] FIG. 6 is a flowchart showing an example of a process (see S3 in FIG. 1) in which the learning device adapts a trained deep learning model to an operational environment. [Figure 7] Figure 7 is a diagram explaining ImageNet-C. [Figure 8] FIG. 8 is a diagram showing the hyperparameters used in the experiment. [Figure 9] FIG. 9 shows the evaluation results of a deep learning model adapted to an operational environment (ImageNet-C) by a learning device. [Figure 10] FIG. 10 is a diagram illustrating an example of the configuration of a computer that executes a learning program. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Hereinafter, a mode (embodiment) for carrying out the present invention will be described with reference to the drawings, but the present invention is not limited to the present embodiment.

[0011] [overview] First, an overview of the learning device of this embodiment will be described with reference to Figures 1 and 2. The learning device adapts a deep learning model to an operating environment by using data of the operating environment of the deep learning model.

[0012] For example, the learning device adapts a deep learning model (hereinafter abbreviated as model) to the operational environment by performing the following two processes. (1) The distribution of feature vectors in the intermediate layer of a deep learning model is made to match as closely as possible between the learning environment and the operational environment of the deep learning model. (2) Correct the confidence level of the value (predicted value) output by the deep learning model for data in the operational environment (operational data).

[0013] As shown in FIG. 1, for example, the learning device calculates statistics on the learning data (data of the learning environment) of a deep learning model (S1). Then, the learning device uses the statistics of the learning data to calculate correlation for each dimension of the feature vector of the learning data, and groups each dimension based on the calculated correlation value (S2: grouping of dimensions based on correlation of feature vectors). After that, the learning device adapts the model to the operating environment using the result of grouping the dimensions of the feature vector in S2 (S3). For example, the learning device performs the above processes (1) and (2).

[0014] Next, an overview of the process executed by the learning device will be described with reference to FIG. 2. In FIG. 2, learning data: {x}, operational data: {x~}, and a trained deep learning model: h(x)=g φ (f θ (x)), feature extractor: f θ , classifier:g φ The parameters θ and φ are defined as follows:

[0015]

number

[0016] For example, the learning device may group the dimensions of the feature vector based on the statistics of the learning data, and then extract the feature vector from the trained deep learning model as shown in Figure 2. θ The learning device obtains the feature vector of the training data {x} and the feature vector of the operational data {x~} from ((1)). The learning device then calculates the degree of discrepancy between the distributions of these feature vectors.

[0017] For example, the learning device can be used to find the group of dimensions of the feature vector (G i) for each feature vector of the training data, we use the statistics (μ Gi ,Σ Gi ) and the statistics showing the distribution of the feature vectors of the operational data (μ~ Gi ,Σ~ Gi ) and the learning device calculates μ Gi ,Σ Gi And, μ~ Gi ,Σ~ Gi Using this, we calculate the degree of inconsistency in the distribution of feature vectors (L a ) to calculate

[0018] The learning device also calculates a loss related to the confidence of the output of the deep learning model in the operational environment ((2)). For example, the learning device calculates a loss related to the confidence of the predicted value output by the deep learning model for the operational data by using information maximization loss (L IM ) to calculate

[0019] Then, the learning device calculates the degree of discrepancy (L a ) and the loss (L IM ) and the feature extractor f θ Update ((3)).

[0020] In this way, the learning device can adapt the deep learning model to the distribution of operational data, and as a result, the learning device can suppress deterioration in the accuracy of the deep learning model even if the distribution of input data during operation of the deep learning model differs from the distribution of input data during training of the deep learning model.

[0021] [Configuration example] Next, a configuration example of the learning device 20 will be described with reference to Fig. 3. The learning device 20 includes an interface unit 21, a storage unit 22, and a control unit 23, for example.

[0022] The interface unit 21 is an interface for inputting and outputting data. For example, the interface unit 21 is a network interface card (NIC). The interface unit 21 can transmit and receive data to and from other devices. The interface unit 21 accepts input of, for example, learning data used in learning a deep learning model and operational data (data to be processed by the deep learning model), which is data in the operational environment of the deep learning model.

[0023] The interface unit 21 may also be connected to input devices such as a mouse, a keyboard, etc. The interface unit 21 may also be connected to output devices such as a display and a speaker.

[0024] The storage unit 22 is a storage device such as a hard disk drive (HDD), a solid state drive (SSD), an optical disk, etc. The storage unit 22 may be a semiconductor memory in which data can be rewritten, such as a random access memory (RAM), a flash memory, or a non-volatile static random access memory (NVSRAM). The storage unit 22 stores an operating system (OS) and various programs executed by the learning device 20, and deep learning model information 211.

[0025] The deep learning model information 221 is parameters of a trained deep learning model. The deep learning model information 221 includes parameters of a feature extractor that extracts a feature vector of input data in the deep learning model, and parameters of a classifier that classifies the input data using the feature vector.

[0026] The control unit 23 controls the entire learning device 20. The control unit 23 is, for example, an electronic circuit such as a CPU (Central Processing Unit), an MPU (Micro Processing Unit), or a GPU (Graphics Processing Unit), or an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array). The control unit 23 also has an internal memory for storing programs that define various processing procedures and control data, and executes each process using the internal memory.

[0027] The control unit 23 functions as the following units by the operation of various programs. For example, the control unit 23 functions as a processing unit 230, a statistics calculation unit 231, a group division unit 232, a first calculation unit 233, a second calculation unit 234, and an update unit 235.

[0028] The processing unit 230 inputs data to a deep learning model constructed based on the deep learning model information 221, and obtains a processing result by the deep learning model. For example, the processing unit 230 inputs image data to a deep learning model, and obtains a recognition result output by the deep learning model. The recognition result is, for example, information specifying a class of an object appearing in the image data, and a score for each class of the object. For example, if objects that may appear in the image data are cats, dogs, monkeys, etc., the recognition result is information such as cats: 80%, dogs: 10%, monkeys: 5%, etc.

[0029] The statistics calculation unit 231 calculates statistics of the learning data used in learning the deep learning model. For example, the statistics calculation unit 231 calculates statistics of the feature extractor f θ training data {x} n i=1 The statistics calculation unit 231 then inputs the feature vectors of the training data and calculates the mean vector μ∈R of the feature vectors based on, for example, equation (1). d´ and the variance-covariance matrix Σ∈R d´×d´ Calculate:

[0030]

number

[0031] The group dividing unit 232 divides each dimension constituting the feature vector of the training data into k groups based on the degree of correlation between the dimensions.

[0032] For example, the group division unit 232 uses the statistics (e.g., variance-covariance matrix Σ) of the feature vectors of the training data calculated by the statistics calculation unit 231 to calculate a correlation matrix C∈R d´×d´ Calculate.

[0033]

number

[0034] Then, the group dividing unit 232 divides the dimensions of the feature vectors into groups having strong correlations with each other, based on the correlation matrix C of the feature vectors of the training data.

[0035] For example, the group division unit 232 divides each dimension {1, ..., d'} of the feature vector into nodes, and the absolute value of the correlation coefficient |C ij Then, the group division unit 232 performs spectral clustering on the graph with edge weights of |. Using the results of the spectral clustering, the group division unit 232 divides each dimension (node) {1, ..., d'} into k groups G i (i∈{1,…,k}) into group G i is expressed by the following equation (3).

[0036]

number

[0037] The first calculation unit 233 calculates the degree of discrepancy between the distribution of a feature vector (first feature vector) of training data of a deep learning model and the distribution of a feature vector (second feature vector) of operational data of the deep learning model.

[0038] For example, the first calculation unit 233 calculates the degree of discrepancy between the distribution of the first feature vector and the distribution of the second feature vector for each group of the dimension of the feature vector divided by the group division unit 232 .

[0039] As one example, the first calculation unit 233 regards the distribution of the feature vectors of the training data and the distribution of the feature vectors of the operational data as multivariate Gaussian distributions, and calculates the degree of discrepancy based on the KL divergence.

[0040] For example, the first calculation unit 233 calculates the KL divergence of the values ​​of each dimension of the feature vectors of the training data and the feature vectors of the operation data for each dimensional group based on the result of grouping the dimensions of the feature vectors by the group division unit 232, and calculates the average value. Then, the first calculation unit 233 regards the calculated average value as the degree of mismatch between the distribution of the feature vectors of the training data and the distribution of the feature vectors of the operation data.

[0041] In addition, the first calculation unit 233 calculates the degree of mismatch for each dimension of the feature vector for each group of the dimensions divided based on the spectral clustering, so that the degree of mismatch can be calculated by focusing on important correlations of each dimension of the feature vector. Details of the process by the first calculation unit 233 will be described later.

[0042] The second calculation unit 234 calculates a loss related to the confidence of the value output by the deep learning model for the operational data. For example, the second calculation unit 234 uses Information Maximization loss (see Non-Patent Document 1) to calculate a loss related to the confidence of the value output by the deep learning model for the operational data.

[0043] For example, the second calculation unit 234 uses a mini-batch of operational data to calculate an information maximization loss regarding the confidence of the predicted probability for each class output by the deep learning model for the operational data, based on the following equation (4).

[0044]

number

[0045] The update unit 235 updates the parameters of the feature extractor so that the sum of the value of the degree of discrepancy in the distribution of the feature vectors calculated by the first calculation unit 233 and the loss in the confidence of the value output by the deep learning model calculated by the second calculation unit 234 is as small as possible.

[0046] For example, the update unit 235 updates the value of the degree of dissimilarity with the distribution of the feature vectors calculated by the first calculation unit 233 (L a ) and the loss (L IM ) is as small as possible, the parameters of the feature extractor are updated using the backpropagation method.

[0047] Thereafter, the processing unit 230 executes image recognition processing and the like on the input data using the deep learning model in which the parameters of the feature extractor have been updated.

[0048] In addition, the update unit 235 updates L IM By using this to update the parameters of the feature extractor of the deep learning model, it is possible to make the deep learning model make predictions for each data set with smaller entropy (higher confidence), while making more diverse predictions overall (higher entropy).

[0049] According to such a learning device 20, even if the distribution of input data during operation of the deep learning model differs from the distribution of input data during training of the deep learning model, deterioration in the accuracy of the deep learning model during operation can be suppressed.

[0050] [Example of processing procedure] Next, an example of the procedure of the process (see S3 in FIG. 1) in which the learning device 20 adapts the trained deep learning model to the operating environment will be described with reference to FIG. 4. It is assumed that the learning device 20 has already performed grouping of each dimension of the feature vector.

[0051] First, the first calculation unit 233 of the learning device 20 samples a mini-batch from the operational data (S31). Then, the first calculation unit 233 calculates the degree of discrepancy L between the distribution of the feature vectors of the sampled mini-batch and the distribution of the feature vectors of the learning data. a (S32: Calculate the degree of discrepancy in the distribution of feature vectors L a calculation).

[0052] In addition, the second calculation unit 234 calculates a loss L related to the confidence of the prediction output of the deep learning model. IM (S33). Then, the update unit 235 calculates L a , L IM We perform backpropagation on the total loss of the feature extractor f θ (S34). If the update unit 235 determines that the termination condition is satisfied (for example, the total loss is equal to or less than a designated value, or the designated number of learning steps is reached) (Yes in S35), the process ends. On the other hand, if the update unit 235 determines that the termination condition is not satisfied (No in S35), the process returns to S31.

[0053] Next, an example of the processing procedure of S32 in FIG. 4 will be described with reference to FIG. 5. The first calculation unit 233 calculates, for example, the operational data (mini-batch) {x i ~} i=1 B The feature extractor f θ and the feature vector {f θ (x i ~)} i=1 B Obtain (S321).

[0054] After S321, the first calculation unit 233 calculates the mean vector μ̂ and the variance-covariance matrix Σ̂ for the feature vector of the mini-batch calculated in S321 (S322). For example, the first calculation unit 233 calculates the mean vector μ̂ and the variance-covariance matrix Σ̂ for the feature vector of the mini-batch based on the following formula (5).

[0055]

number

[0056] After S322, the first calculation unit 233 calculates the group G from the mean vector μ~ and the variance-covariance matrix Σ~. i The first calculation unit 233 extracts the rows and columns corresponding to the group G from the mean vector μ and the variance-covariance matrix Σ calculated in advance using the training data (S323). i The rows and columns corresponding to the dimensions included in are extracted (S324).

[0057] After that, the first calculation unit 233 calculates the matrix extracted in S323 and the matrix extracted in S324 using two multivariate normal distributions N(μ Gi ,Σ~ Gi ), N(μ Gi ,Σ Gi ) and calculate the discrepancy L based on the KL divergence. a For example, the first calculation unit 233 calculates the degree of mismatch L by the following formula (6): a Calculate.

[0058]

number

[0059] In addition, when the size B of the operational data (mini-batch) is smaller than the number of dimensions (d') of the feature vector, |Σ~| in the above formula (6) becomes 0. Therefore, the first calculation unit 233 calculates the degree of discrepancy L of the distribution of the feature vector. aTherefore, the learning device 20 solves the above problem by using the group division unit 232 to group each dimension of the feature vector and reduce the number of dimensions of the feature vector.

[0060] According to the above-described learning device 20, a trained deep learning model can be adapted to an operating environment. This makes it possible to suppress a decrease in accuracy of the deep learning model even if the distribution of input data during operation of the deep learning model differs from the distribution of input data during training of the deep learning model.

[0061] [Other embodiments] Even if the learning device 20 performs grouping for each dimension of the feature vector, the variance-covariance matrix Σ Gi is degenerate, so |Σ Gi |=0, and the degree of inconsistency in the distribution of the feature vectors L a In such a case, the learning device 20 performs, for example, the following process.

[0062] For example, as shown in Fig. 6, the statistics calculation unit 231 calculates statistics regarding the training data (S331) in the same manner as in S1 of Fig. 1. For example, the statistics calculation unit 231 calculates a mean vector μ∈R d´ and the variance-covariance matrix Σ∈R d´×d´ Calculate:

[0063] 1, the group division unit 232 performs grouping of dimensions based on the correlation of feature vectors (S332). For example, the group division unit 232 calculates a correlation matrix C of the feature vectors of the training data using the variance-covariance matrix Σ of the feature vectors of the training data, and divides the dimensions of the feature vectors into a group G having strong correlations with each other. i Divide into:

[0064] Then, the first calculation unit 233 calculates |Σ Gi Calculate | and |Σ Gi If it is determined that |=0 (Yes in S333), the variance-covariance matrix ΣGi In order to identify the direction in which the variance-covariance matrix Σ Gi Then, the process proceeds to S335. Note that this vector is calculated using the variance-covariance matrix Σ Gi The eigenvalue represents the variance in the direction of the corresponding eigenvector (principal component). Gi If it is determined that |=0 (No in S333), the process from S323 onward in FIG. 5 is executed.

[0065]

number

[0066] In S335 of FIG. 6, the first calculation unit 233 calculates whether the eigenvalue of the vector is greater than a predetermined threshold ε (for example, 10 -5 If it is determined that the eigenvalue λ is smaller than the threshold value ε (Yes in S335), the corresponding principal component direction is regarded as degenerate, and a correction is performed to replace the eigenvalue with a threshold value ε (S336). For example, the first calculation unit 233 calculates the eigenvalue λ as shown in formula (8). Gi j Then, the process proceeds to S337. On the other hand, the first calculation unit 233 performs a correction to replace the eigenvalue of the vector with a predetermined threshold ε (for example, 10 -5 If it is determined that the number of the reads is equal to or greater than the predetermined number (No in S335), the process skips S336 and proceeds to S337.

[0067]

number

[0068] In S337, the first calculation unit 233 calculates the variance-covariance matrix Σ Gi Principal component V Gi In the space spanned by Gi , Σ Gi , μ~ Gi , Σ~ Gi Map the following.

[0069]

number

[0070] After S337, the first calculation unit 233 calculates the variance-covariance matrix Σ Gi Principal component V Gi A new μ Gi , Σ Gi , μ~ Gi , Σ~ Gi Using the discrepancy L a For example, the first calculation unit 233 calculates μ Gi , Σ Gi , μ~ Gi , Σ~ Gi Substituting this into equation (6) gives the discrepancy L a Calculate.

[0071] In this way, even if the learning device 20 performs grouping of each dimension of the feature vector, |Σ Gi Even if |=0, the degree of inconsistency in the distribution of feature vectors L a can be calculated.

[0072] Although the second calculation unit 234 of the learning device 20 calculates the loss related to the confidence of the value output by the deep learning model for the operational data based on the formula (4), the calculation is not limited to this. For example, the second calculation unit 234 may perform the calculation using only the value of the second term of the formula (4).

[0073] Furthermore, the learning device 20 divides each dimension of the feature vector into a plurality of groups by the group division unit 232, but is not limited to this. For example, when the number of dimensions (d') of the feature vector is smaller than the size B of the feature operation data (mini-batch), the learning device 20 may not divide each dimension of the feature vector.

[0074] [Evaluation experiment] Next, the results of an evaluation experiment of a deep learning model adapted to an operational environment by the learning device 20 will be described. Here, the learning device 20 adapted a deep learning model that had been trained to classify ImageNet to a noisy dataset (ImageNet-C (see Non-Patent Document 2)). ImageNet-C is a dataset in which various types of noise (corruption) are added to the images of ImageNet, as shown in FIG. 7.

[0075] The learning device 20 adapted the trained deep learning model to ImageNet-C by the following steps (1) to (4).

[0076] (1) Obtaining a deep learning model trained on ImageNet The learning device 20 acquired a trained model (ResNet-50) of the Torchvision library as the deep learning model to be adapted.

[0077] (2) Calculating statistics for training data (ImageNet) The learning device 20 uses the intermediate layer output (d' = 2048 dimensions) of the above-mentioned ResNet-50 as a feature vector, and calculates the mean vector μ ∈ R 2048 and the variance-covariance matrix Σ∈R 2048×2048 was calculated.

[0078] (3) Dimension grouping based on correlation of feature vectors The learning device 20 divided the 2048 dimensions of the feature vectors of the training data into 64 groups by spectral clustering.

[0079] (4) Adapting the deep learning model to the production environment (ImageNet-C) The learning device 20 adapted the deep learning model to the operational environment (ImageNet-C) with the hyperparameters shown in FIG. 8 .

[0080] [Evaluation Results] The evaluation results of the accuracy of the deep learning model adapted to the operational environment (ImageNet-C) by the learning device 20 are shown in Fig. 9. Here, the accuracy (correct answer rate) of the deep learning model adapted to the operational environment (ImageNet-C) was evaluated for each type of corruption.

[0081] 9, before adaptation is a result of predicting ImageNet-C data using a deep learning model trained on ImageNet as is, while after adaptation is a result of predicting ImageNet-C data using a deep learning model adapted to ImageNet-C by learning device 20.

[0082] As shown in FIG. 9, it was confirmed that the deep learning model adapted to ImageNet-C by the learning device 20 improved accuracy (correct answer rate) for all types of corruption.

[0083] [System configuration, etc.] In addition, each component of each unit shown in the figure is a functional concept, and does not necessarily have to be physically configured as shown in the figure. In other words, the specific form of distribution and integration of each device is not limited to that shown in the figure, and all or a part of it can be functionally or physically distributed and integrated in any unit according to various loads, usage conditions, etc. Furthermore, each processing function performed by each device can be realized in whole or in part by a CPU and a program executed by the CPU, or can be realized as hardware using wired logic.

[0084] Furthermore, among the processes described in the above embodiments, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically by a known method. In addition, the information including the processing procedures, control procedures, specific names, various data and parameters shown in the above documents and drawings can be changed arbitrarily unless otherwise specified.

[0085] [program] The learning device 20 can be implemented by installing a program (learning program) as package software or online software on a desired computer. For example, by executing the above program on an information processing device, the information processing device can function as the learning device 20. The information processing device referred to here includes mobile communication terminals such as smartphones, mobile phones, and PHS (Personal Handyphone Systems), as well as terminals such as PDAs (Personal Digital Assistants).

[0086] 10 is a diagram showing an example of a computer that executes a learning program. The computer 1000 has, for example, a memory 1010 and a CPU 1020. The computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. Each of these components is connected by a bus 1080.

[0087] The memory 1010 includes a read only memory (ROM) 1011 and a random access memory (RAM) 1012. The ROM 1011 stores a boot program such as a basic input output system (BIOS). The hard disk drive interface 1030 is connected to a hard disk drive 1090. The disk drive interface 1040 is connected to a disk drive 1100. A removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to a mouse 1110 and a keyboard 1120, for example. The video adapter 1060 is connected to a display 1130, for example.

[0088] The hard disk drive 1090 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, the program that defines each process executed by the above-mentioned learning device 20 is implemented as a program module 1093 in which computer-executable code is written. The program module 1093 is stored, for example, in the hard disk drive 1090. For example, a program module 1093 for executing a process similar to the functional configuration of the learning device 20 is stored in the hard disk drive 1090. The hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

[0089] Data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, the memory 1010 or the hard disk drive 1090. Then, the CPU 1020 reads out the program module 1093 or the program data 1094 stored in the memory 1010 or the hard disk drive 1090 into the RAM 1012 as necessary and executes them.

[0090] Note that the program module 1093 and the program data 1094 are not limited to being stored in the hard disk drive 1090, but may be stored in, for example, a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and the program data 1094 may be stored in another computer connected via a network (such as a local area network (LAN) or wide area network (WAN)). The program module 1093 and the program data 1094 may then be read by the CPU 1020 from the other computer via the network interface 1070. [Explanation of symbols]

[0091] 20 Learning Device 21 Interface section 22 Memory section 23 Control Unit 230 Processing section 231 Statistics calculation section 232 Group division section 233 First Calculation Section 234 Second Calculation Section 235 Update Department

Claims

1. a first calculation unit that calculates a degree of discrepancy between a distribution of a first feature vector, which is a feature vector in an intermediate layer of a deep learning model of learning data used for learning the deep learning model, and a distribution of a second feature vector, which is a feature vector in an intermediate layer of the deep learning model of data to be processed; A second calculation unit that calculates a loss related to a confidence level of a value output by the deep learning model for the processing target data; an update unit that updates parameters of a feature extractor that extracts feature vectors in the deep learning model so that the sum of the discrepancy value and the loss related to the confidence is as small as possible; A learning device comprising:

2. a group division unit that divides each of the dimensions constituting the first feature vector into a plurality of groups based on a correlation between the dimensions, The first calculation unit: Calculating a degree of mismatch between the distribution of the first feature vector and the distribution of the second feature vector by calculating a degree of mismatch between each dimension value of the first feature vector and each dimension value of the second feature vector for each group.

2. The learning device according to claim 1 .

3. The first calculation unit: The distribution of the first feature vector and the distribution of the second feature vector are regarded as multivariate Gaussian distributions, and the degree of discrepancy between the distribution of the first feature vector and the distribution of the second feature vector is calculated by KL divergence.

2. The learning device according to claim 1 .

4. The first calculation unit: For each group, the KL divergence of the values ​​of each dimension of the first feature vector and the second feature vector is calculated to calculate the degree of discrepancy between the distribution of the first feature vector and the distribution of the second feature vector.

3. The learning device according to claim 2.

5. The second calculation unit is Information Maximization loss is used to calculate the loss related to the confidence of the value output by the deep learning model for the processing target data.

2. The learning device according to claim 1 .

6. A learning method executed by a learning device, comprising: Calculating a degree of discrepancy between a distribution of a first feature vector, which is a feature vector in an intermediate layer of a deep learning model of training data used for training the deep learning model, and a distribution of a second feature vector, which is a feature vector in an intermediate layer of the deep learning model of data to be processed; Calculating a loss related to the confidence of a value output by the deep learning model for the processing target data; updating parameters of a feature extractor that extracts feature vectors in the deep learning model so that the sum of the discrepancy value and the loss related to the confidence is as small as possible; A learning method comprising:

7. Calculating a degree of discrepancy between a distribution of a first feature vector, which is a feature vector in an intermediate layer of a deep learning model of training data used for training the deep learning model, and a distribution of a second feature vector, which is a feature vector in an intermediate layer of the deep learning model of data to be processed; Calculating a loss related to the confidence of a value output by the deep learning model for the processing target data; updating parameters of a feature extractor that extracts feature vectors in the deep learning model so that the sum of the discrepancy value and the loss related to the confidence is as small as possible; A learning program for a computer to execute the above.

Images