Adaptation device and adaptation method

Abstract

This adaptation device inputs training data to a trained deep learning model and calculates a distribution of a first feature vector which is a feature vector output from an intermediate layer of the deep learning model. Further, the adaptation device inputs data pertaining to an operation environment to the trained deep learning model and calculates a distribution of a second feature vector which is a feature vector output from the intermediate layer of the deep learning model. Then, the adaptation device updates the parameter of the trained deep learning model so as to bring the distribution of the second feature vector close to the distribution of the first feature vector.

Description

Adaptation device and adaptation method

[0001] The present invention relates to an adaptation device and an adaptation method for adapting a deep learning model to an operational environment.

[0002] In deep learning, if the distribution of input data for a model changes between when the model is trained and when it is put into operation, the classification accuracy of the model may decline. To prevent this decline in classification accuracy, there is a technology that adapts a trained classification model to the operating environment through unsupervised learning using data from the operating environment of the classification model.

[0003] This technology adapts a classification model to the operating environment by, for example, updating the parameters of the classification model to minimize the uncertainty (entropy) of the predicted output (predicted probability of each class) of the classification model in the operating environment.

[0004] Wang Dequan, et al., Fully Test-time Adaptation by Entropy Minimization, International Conference on Learning Representations. 2020.

[0005] However, the above-mentioned conventional technology cannot be applied to models other than classification models. For example, in the case of a regression model, only one predicted value is output, so entropy cannot be calculated. Therefore, with the conventional technology, it has been difficult to adapt models other than classification models to the operating environment. Therefore, an object of the present invention is to adapt models other than classification models to the operating environment.

[0006] In order to solve the above-mentioned problems, the present invention is characterized by comprising a first calculation unit that inputs training data of a deep learning model into the trained deep learning model and calculates the distribution of a first feature vector, which is a feature vector output from an intermediate layer of the deep learning model; a second calculation unit that inputs data of the operating environment of the deep learning model into the trained deep learning model and calculates the distribution of a second feature vector, which is a feature vector output from the intermediate layer of the deep learning model; and a model update unit that updates the parameters of the trained deep learning model so that the distribution of the second feature vector approaches the distribution of the first feature vector.

[0007] According to the present invention, models other than classification models can also be adapted to the operating environment.

[0008] FIG. 1 is a diagram showing an example of the configuration of an adaptive system. FIG. 2 is a flowchart showing an example of a processing procedure executed by a learning device of the adaptive system. FIG. 3 is a flowchart showing an example of a processing procedure executed by a model adaptation device of the adaptive system. FIG. 4 is a diagram showing a training dataset of a regression model. FIG. 5 is a diagram showing a training dataset and an operating environment dataset of a regression model. FIG. 6 is a diagram showing an evaluation result of the predictive performance of a regression model to which the adaptive system is applied. FIG. 7 is a diagram showing an evaluation result of the predictive performance of a regression model to which the adaptive system is applied. FIG. 8 is a diagram showing an example of a computer that executes an adaptation program.

[0009] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a description will be given of an embodiment of the present invention with reference to the drawings, but the present invention is not limited to the embodiment.

[0010] [Overview] First, an overview of the adaptation device (adaptation system) of this embodiment will be described. The adaptation system adapts a deep learning model trained in a learning environment to the operational environment of the deep learning model. For example, the adaptation system updates the parameters of the trained deep learning model so that the distribution of feature vectors output by the intermediate layer of the deep learning model in the operational environment approaches the distribution of feature vectors output by the intermediate layer of the deep learning model in the learning environment. In this way, the adaptation system can adapt models other than classification models to the operational environment.

[0011] [Configuration Example] Next, a configuration example of an adaptation system will be described using Fig. 1. The adaptation system includes a learning device 10 and a model adaptation device 20. The learning device 10 uses training data to train a deep learning model. The learning device 10 includes a learning unit 11 and a first statistics calculation unit 12.

[0012] The learning unit 11 learns a deep learning model using training data. The first statistics calculation unit 12 calculates statistics (feature statistics) of feature vectors output by intermediate layers of the trained deep learning model. For example, the first statistics calculation unit 12 inputs training data to the trained deep learning model and calculates the distribution of feature vectors (first feature vectors) output by intermediate layers of the deep learning model.

[0013] The model adaptation device 20 adapts the deep learning model learned by the learning device 10 to an operational environment. The model adaptation device 20 includes a data acquisition unit 21, a second statistics calculation unit 22, and a model update unit 23.

[0014] The data acquisition unit 21 acquires data on the operational environment of the deep learning model. The second statistics calculation unit 22 calculates statistics (feature statistics) of feature vectors output by intermediate layers of the deep learning model in the operational environment. For example, the second statistics calculation unit 22 inputs the operational environment data to the deep learning model and calculates the distribution of feature vectors (second feature vectors) output from the intermediate layers of the deep learning model.

[0015] The model update unit 23 updates the parameters of the trained deep learning model so that the distribution of the second feature vector approaches the distribution of the first feature vector.

[0016] Here, the model updating unit 23 performs, for example, principal component analysis of the elements constituting the first feature vector and the second feature vector, and extracts a predetermined number of elements in descending order of contribution. Then, the model updating unit 23 updates the parameters of the deep learning model so that the distribution of the second feature vector in the subspace constituted by the extracted elements approaches the distribution of the first feature vector. In this way, the model updating unit 23 can bring the distributions in a particularly important subspace of the feature space closer together.

[0017] The parameters of the trained or updated deep learning model and the feature statistics of the deep learning model calculated by the first statistics calculation unit 12 are stored in the memory unit of the adaptive system.

[0018] The adaptive system, learning device 10, and model adaptation device 20 are realized by a computer including, for example, an input / output interface, a CPU (Central Processing Unit), and a storage unit. The functions of the learning unit 11, first statistics calculation unit 12, data acquisition unit 21, second statistics calculation unit 22, and model update unit 23 are realized by, for example, a CPU included in the adaptive system executing a program stored in the storage unit.

[0019] 2 and 3, an example of a processing procedure executed by the adaptive system will be described. First, the learning unit 11 of the learning device 10 learns a deep learning model using training data (S1). Then, the first statistics calculation unit 12 inputs the training data to the deep learning model trained in S1 and calculates statistics (feature statistics) of feature vectors output from the intermediate layer of the deep learning model (S2).

[0020] After S2, the data acquisition unit 21 of the model adaptation device 20 acquires data on the operating environment of the deep learning model. Then, the second statistics calculation unit 22 inputs the data on the operating environment to the trained deep learning model and calculates statistics (feature statistics) of the feature vectors output from the intermediate layer of the deep learning model (S11 in FIG. 3).

[0021] After S11, the model update unit 23 updates the parameters of the deep learning model so that the feature statistics calculated in S11 approach the feature statistics calculated in S2 of Fig. 2 (S12: model update). The model adaptation device 20 then uses the deep learning model with updated parameters to output prediction results for data in the operational environment (S13). The process then returns to S11. The model adaptation device 20 adapts the trained deep learning model to the operational environment by repeating the above-described processes from S11 to S13.

[0022] [Details of Processing] Next, the processing of FIGS. 2 and 3 will be described in detail.

[0023] [S1: Learning] In S1 of FIG. 2, the learning unit 11 uses a learning data set shown in the following equation 1 to learn a deep learning model f θ For the learning, any conventional learning algorithm and loss function may be used.

[0024]

[0025] [S2: Statistical Calculation] In S2, the first statistical calculation unit 12 calculates, for example, the statistics of the deep learning model f θ The hidden feature vector z of the training dataset D S Statistics S (D S ) is calculated.

[0026] In addition, the deep learning model f θ is the feature extractor z=g φ (x) and predictor y=h ψ (z) (y=f θ (x)=h ψ (g φ (x), θ=(φ,ψ)). Deep learning model f θ can be divided into any number of intermediate layers.

[0027] In addition, the statistic S (D S ) are, for example, the mean of the feature vector z, the variance / covariance matrix, and the principal components (eigenvalues / eigenvectors of the variance / covariance matrix).

[0028] The first statistics calculation unit 12 calculates the learning data set D of the feature vector z based on, for example, the following equation (1): S Calculate the mean μ for

[0029]

[0030] Furthermore, the first statistics calculation unit 12 calculates the learning data set D of the feature vector z based on, for example, the following equation (2): S Calculate the variance-covariance matrix for

[0031]

[0032] Furthermore, for example, the first statistics calculation unit 12 calculates the principal components (eigenvalues ​​and eigenvectors of the variance and covariance matrices) of the feature vector z shown in the following equation 4.

[0033]

[0034] [S11: Statistical Calculation (During Model Adaptation)] In S11 of FIG. 3, the second statistical calculation unit 22 calculates, for example, data (mini-batch) B={x i} i=1 B For the deep learning model f θ For example, the second statistics calculation unit 22 calculates the statistics S(B) of the feature vector z of the hidden layer of the feature extractor z=g φ Using (x), calculate the statistic S(B) in the same way as S2 above.

[0035] [S12: Model Update] In S12, the model update unit 23 performs a process of updating the model by, for example, determining whether the statistics S(B) of the data in the operational environment are equal to the statistics S(D) of the learning dataset. S ) and update the parameters of the deep learning model to get closer to

[0036] For example, in advance, the user of the adaptive system may select the statistics S(B) and S(DS ) loss function d(S(B),S(D S ) are defined in advance. Then, the model update unit 23 updates all or part of the parameters θ (for example, only φ) of the deep learning model by a gradient method using the loss function d. For example, the Kullback-Leibler divergence, Euclidean distance, or Wasserstein distance is used as the loss function d.

[0037] [S13: Prediction] In S13, the adaptive system predicts y for the data of the operational environment using the deep learning model whose parameters have been updated in S12 (see equation (3) below).

[0038]

[0039] [Evaluation Results] Next, the evaluation results when the adaptive system of this embodiment is applied to a regression model will be described.

[0040] [Learning Environment] First, we will explain the model training. As shown in Figure 4, the adaptive system used SVHN as training data and trained ResNet-26 (a model that predicts numbers in images). The adaptive system also used UTKFace as training data and trained ResNet-50 (a model that predicts the age of a person in an image).

[0041] SVHN: The Street View House Numbers (SVHN) Dataset (stanford.edu) UTKFace: Large Scale Face Dataset (susanqq.github.io)

[0042] [Model Adaptation] The adaptation system also used the output (2048 dimensions) of the layer just before the final layer of each of the above models as a feature vector. The adaptation system also performed principal component analysis of the elements (components) that make up the feature vector, and calculated the contribution of each component. The adaptation system then calculated the feature vector statistics by calculating the mean μ of the feature vector and the eigenvalues ​​λ1,…,λ of the top 100 components in order of contribution. 100 ・Eigenvectors v1,…,v 100 and was used.

[0043] The adaptive system calculates the feature vectors of the data in the operational environment of each model as statistics of the training data and converts them into eigenvectors v1,…,v 100 (See equation (4) below.) In equation (4) below, z' is the projected feature vector.

[0044]

[0045] The adaptive system calculates the mean μ B and variance σ 2 is the statistic S(B). The adaptive system calculates the mean μ B is calculated by the following formula (5), and the variance σ 2 was calculated using the following formula (6).

[0046]

[0047] The adaptive system then calculates the Kullback-Leibler information of the following two normal distributions as the loss function d(S(B),S(D S The model parameters were updated to minimize

[0048] ・Normal distribution N(μ B ,σ 2 ) ・Normal distribution N(0,λ), λ=(λ1,…,λ 100 )

[0049] Loss function d(S(B),S(D S ) is shown in the following equation (7).

[0050]

[0051] [Operational Environment] The adaptive system used the dataset shown in Figure 5 as the operational environment data for the model. That is, the adaptive system adapted a deep learning model trained with SVHN to MNIST. The adaptive system also adapted a deep learning model trained with UTKFace to UTKFace (with noise). UTKFace (with noise) is data obtained by applying noise to UTKFace. In this experiment, 13 types of noise shown in Figure 7 were applied to UTKFace.

[0052] [Evaluation Results] The evaluation results of the prediction performance of the regression model adapted to the operating environment by the application system of this embodiment are shown in Figs. 6 and 7. The comparative example is the regression model before adaptation to the operating environment. The evaluation index is R 2 was used.

[0053]

[0054] As shown in FIG. 6, when the regression model trained with SVHN was applied to MNIST using the application system of this embodiment, it was confirmed that the prediction performance was improved compared to the comparative example.

[0055] Furthermore, as shown in Figure 7, it was confirmed that when a regression model trained with UTKFace was adapted to UTKFace (with noise), the prediction performance was improved compared to the comparative example.

[0056] From the above, it has been confirmed that the adaptation system of this embodiment can also adapt models other than classification models to the operating environment.

[0057] [System Configuration, etc.] The components of each unit shown in the figure are conceptual functional units and do not necessarily have to be physically configured as shown. In other words, the specific form of distribution and integration of each device is not limited to that shown, and all or part of them can be functionally or physically distributed and integrated in any unit depending on various loads, usage conditions, etc. Furthermore, all or any part of the processing functions performed by each device can be realized by a CPU and a program executed by the CPU, or can be realized as hardware using wired logic.

[0058] 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 using 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 as desired unless otherwise specified.

[0059] [Program] The adaptive system described above can be implemented by installing a program (adaptive program) as package software or online software on a desired computer. For example, by executing the program on an information processing device, the information processing device can function as the adaptive system. 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).

[0060] 8 is a diagram showing an example of a computer that executes an adaptation program. The computer 1000 includes, for example, a memory 1010 and a CPU 1020. The computer 1000 also includes 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. These components are connected by a bus 1080.

[0061] 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 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.

[0062] 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 adaptive system 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 processes similar to those of the functional configuration of the adaptive system is stored in the hard disk drive 1090. The hard disk drive 1090 may be replaced with an SSD (Solid State Drive).

[0063] Data used in the processing of the above-described embodiment is stored as program data 1094, for example, in the memory 1010 or the hard disk drive 1090. The CPU 1020 then reads 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.

[0064] The program module 1093 and program data 1094 may not necessarily be stored in the hard disk drive 1090, but may also be stored in a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (such as a local area network (LAN) or a wide area network (WAN)). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via the network interface 1070.

[0065] REFERENCE SIGNS LIST 10 Learning device 11 Learning unit 12 First statistics calculation unit 20 Model adaptation device 21 Data acquisition unit 22 Second statistics calculation unit 23 Model update unit

Claims

1. An adaptive device comprising: a first calculation unit that inputs training data of a deep learning model into the trained deep learning model and calculates the distribution of a first feature vector, which is a feature vector output from an intermediate layer of the deep learning model; a second calculation unit that inputs data of the operating environment of the deep learning model into the trained deep learning model and calculates the distribution of a second feature vector, which is a feature vector output from an intermediate layer of the deep learning model; and a model update unit that updates parameters of the trained deep learning model so that the distribution of the second feature vector approaches the distribution of the first feature vector.

2. The adaptive device described in claim 1, characterized in that the model update unit extracts a predetermined number of elements from among the elements constituting the first feature vector and the second feature vector in descending order of contribution through principal component analysis, and updates the parameters of the deep learning model after learning so that the distribution of the second feature vector in a subspace constituted by the extracted elements approaches the distribution of the first feature vector.

3. The adaptive device according to claim 1, wherein the deep learning model is a regression model.

4. An adaptation method executed by an adaptation device, comprising: a step of inputting training data of a deep learning model into the trained deep learning model and calculating the distribution of a first feature vector, which is a feature vector output from an intermediate layer of the deep learning model; a step of inputting data of the operating environment of the deep learning model into the trained deep learning model and calculating the distribution of a second feature vector, which is a feature vector output from an intermediate layer of the deep learning model; and a step of updating parameters of the trained deep learning model so that the distribution of the second feature vector approaches the distribution of the first feature vector.

Images