Learning methods, learning devices, and programs

Abstract

This learning method employing batch learning performed by a computer includes: acquiring learning data including first time-series data and second time-series data different from the first time-series data; performing first learning processing to train a neural process model that uses a stochastic process to output a predicted result taking into account uncertainty, such that a first time-series data distribution and a second time-series data distribution are predicted from the first time-series data and the second time-series data; and performing second learning processing employing a contrastive learning algorithm to cause approaching-type learning to be performed between a plurality of items of first sampling data generated by sampling from the first time-series data distribution, to cause approaching-type learning to be performed between a plurality of items of second sampling data generated by sampling from the second time-series data distribution, and to cause separation-type learning to be performed between the plurality of items of first sampling data and the plurality of items of second sampling data.

Description

[Technical Field]

[0001] This disclosure relates to a learning method, a learning device, and a program. [Background technology]

[0002] In AI development, it is necessary to collect a large amount of labeled data in order to obtain a highly accurate model.

[0003] However, even when data can be collected, the cost of labeling it is high, hindering its business application.

[0004] Therefore, there is a need to develop a learning method that can acquire highly accurate models while reducing the amount of labeling required in AI development.

[0005] In contrast, a technique has been disclosed (for example, Non-Patent Document 1) that enables the acquisition of highly accurate models from image data with only about 1% of the labels, using self-supervised learning with data augmentation and comparative learning. Specifically, Non-Patent Document 1 augments the input image data to generate two pairs of versions of the image data. Then, comparative learning is performed to maximize the features of the same image data (making them close together) while minimizing the features of the different image data (making them far apart). In this way, by utilizing data augmentation and comparative learning, it is possible to train a highly accurate model from a small amount of data and a small number of labels. [Prior art documents] [Non-patent literature]

[0006] [Non-Patent Document 1] J. Gordon, WP Bruinsma, AYK Foong, J. Requeima, Y. Dubois, and RE Turner. Convolutional conditional neural processes. In International Conference on Learning Representations (ICLR),2020. [Non-Patent Document 2] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual representations. arXiv:2002.05709, 2020. [Overview of the project] [Problems that the invention aims to solve]

[0007] However, while the technology disclosed in Non-Patent Document 1 above is an effective method for handling image data, it cannot handle time-series data.

[0008] This disclosure is made in light of the circumstances described above, and aims to provide a learning method that can handle time-series data in self-supervised learning. [Means for solving the problem]

[0009] A learning method according to one aspect of the present disclosure is a learning method by batch learning performed by a computer, which involves acquiring training data including first time series data and second time series data different from the first time series data, and performing a first learning process to train a neural process model, which is a deep learning model that outputs prediction results considering uncertainty using a stochastic process, so as to predict a first time series data distribution that shows the statistical properties of the first time series data and a second time series data distribution that shows the statistical properties of the second time series data from the first time series data and the second time series data, and performing a second learning process using a contrast learning algorithm to perform learning that brings together multiple first sampled data generated by sampling from the first time series data distribution as positive samples, learns that brings together multiple second sampled data generated by sampling from the second time series data distribution as positive samples, and learns that moves apart the multiple first sampled data and the multiple second sampled data as negative samples.

[0010] These general or specific embodiments may be implemented as systems, devices, methods, integrated circuits, computer programs, or recording media such as computer-readable CD-ROMs, or as any combination of systems, devices, methods, integrated circuits, computer programs, and recording media. [Effects of the Invention]

[0011] According to this disclosure, it is possible to realize a learning method that can handle time-series data in self-supervised learning. [Brief explanation of the drawing]

[0012] [Figure 1] Figure 1 is a block diagram showing an example of the configuration of a learning device according to an embodiment. [Figure 2] Figure 2 is a conceptual diagram illustrating the processing of the learning device according to the embodiment. [Figure 3] Figure 3 is a conceptual diagram illustrating the processing of the learning device according to the embodiment. [Figure 4A] Figure 4A is a diagram that conceptually illustrates self-supervised contrastive learning. [Figure 4B] Figure 4B is a diagram that conceptually illustrates an example of boundary generation by performing self-supervised contrastive learning as shown in Figure 4A. [Figure 5A] Figure 5A is a diagram that conceptually explains the predictive distribution by neural processes. [Figure 5B] Figure 5B shows an example of the prediction results obtained by the neural process shown in Figure 5A. [Figure 6] Figure 6 is a diagram for conceptually explaining the first learning process and the second learning process according to the embodiment. [Figure 7] Figure 7 shows pseudocode of the algorithm, which is the processing procedure of the learning device according to the embodiment. [Figure 8] Figure 8 is a flowchart illustrating the overview of the operation of the learning device according to the embodiment. [Figure 9] Figure 9 shows the results of evaluating the performance of the model of this disclosure using the dataset related to the experimental example. [Figure 10] This figure shows the accuracy when ContrNP(ours) is trained with different label rates for the AFDB dataset related to the experimental examples. [Modes for carrying out the invention]

[0013] A learning method according to one aspect of the present disclosure is a learning method by batch learning performed by a computer, which involves acquiring training data including first time series data and second time series data different from the first time series data, and performing a first learning process to train a neural process model, which is a deep learning model that outputs prediction results considering uncertainty using a stochastic process, so as to predict a first time series data distribution that shows the statistical properties of the first time series data and a second time series data distribution that shows the statistical properties of the second time series data from the first time series data and the second time series data, and performing a second learning process using a contrast learning algorithm to perform learning that brings together multiple first sampled data generated by sampling from the first time series data distribution as positive samples, learns that brings together multiple second sampled data generated by sampling from the second time series data distribution as positive samples, and learns that moves apart the multiple first sampled data and the multiple second sampled data as negative samples.

[0014] According to this, by combining the framework of self-supervised controlled learning with the learning framework of neural process models, it is possible to realize learning methods that can handle time-series data in self-supervised learning.

[0015] Here, for example, the first time series data may be time series sampled data obtained by sampling time-sequentially continuous first data, and the second time series data may be time series sampled data obtained by sampling time-sequentially continuous second data.

[0016] Furthermore, for example, the first learning process and the second learning process may be performed simultaneously, and when performing the first learning process and the second learning process, an error function may be used which is obtained by modifying the second error function by adding a term of the first error function used in the contrast learning algorithm, which reduces the error in the case of positive samples and increases the error in the case of negative samples, to a term of the second error function related to the error of the prediction result used in the neural process model.

[0017] In this way, the first and second learning processes are performed using an error function that combines the first error function used in the contrastive learning algorithm and the second error function related to the error of the prediction result used in the neural process model. This allows the first learning process, which learns the prediction distribution by the neural process, and the second learning process, which performs feature representation learning using self-supervised contrastive learning, to be performed simultaneously on the target neural process model.

[0018] A learning device according to one aspect of the present disclosure is a learning device for performing learning by batch learning, comprising: an acquisition unit that acquires learning data including first time series data and second time series data different from the first time series data; a learning processing unit that performs a first learning process to train a neural process model, which is a deep learning model that outputs prediction results considering uncertainty using a stochastic process, so as to predict a first time series data distribution that shows the statistical properties of the first time series data and a second time series data distribution that shows the statistical properties of the second time series data from the first time series data and the second time series data; and a learning processing unit that performs a second learning process using a contrast learning algorithm to bring together multiple first sampled data generated by sampling from the first time series data distribution as positive samples, bring together multiple second sampled data generated by sampling from the second time series data distribution as positive samples, and move apart the multiple first sampled data and the multiple second sampled data as negative samples.

[0019] Furthermore, a program according to one aspect of this disclosure is a program that causes a computer to execute a batch learning method, and which causes the computer to execute the following: acquire training data including first time series data and second time series data different from the first time series data; train a neural process model, which is a deep learning model that outputs prediction results that take uncertainty into account using a stochastic process, so as to predict a first time series data distribution that shows the statistical properties of the first time series data and a second time series data distribution that shows the statistical properties of the second time series data from the first time series data and the second time series data; and perform a second learning process that uses a contrast learning algorithm to bring together multiple first sampled data generated by sampling from the first time series data distribution as positive samples, bring together multiple second sampled data generated by sampling from the second time series data distribution as positive samples, and move apart the multiple first sampled data and the multiple second sampled data as negative samples.

[0020] These comprehensive or specific embodiments may be implemented as systems, devices, methods, integrated circuits, computer programs, or recording media such as computer-readable CD-ROMs, or as any combination of systems, devices, methods, integrated circuits, computer programs, and recording media.

[0021] The embodiments of this disclosure will be described below with reference to the drawings. Each embodiment described below is a specific example of this disclosure. The numerical values, shapes, components, steps, and order of steps shown in the following embodiments are examples only and are not intended to limit this disclosure. Furthermore, any components in the following embodiments that are not described in an independent claim will be described as optional components. Also, in all embodiments, the contents of each can be combined.

[0022] (Embodiment) The following will explain the learning method and other aspects related to this embodiment, with reference to the drawings.

[0023] [1 Learning device 1] Figure 1 is a block diagram showing an example of the configuration of the learning device 1 according to this embodiment. Figures 2 and 3 are diagrams conceptually illustrating the processing of the learning device 1 according to this embodiment. Figure 3 is a different representation of the processing in Figure 2.

[0024] Learning device 1 is a device for learning time-series representations as feature representations by utilizing self-supervised learning that combines comparative learning and learning of neural process models.

[0025] In this embodiment, the learning device 1 comprises an acquisition unit 11, an NP model 12, and a learning processing unit 13, as shown in Figure 1.

[0026] [1.1 Acquisition part 11] The acquisition unit 11 includes, for example, a computer including memory and a processor (microprocessor), and various functions are realized by the processor executing a control program stored in memory. Specifically, the acquisition unit 11 acquires training data which includes first time series data and second time series data which is different from the first time series data. Here, for example, the first time series data is time series sampled data obtained by sampling first data which is continuous in time. The second time series data is time series sampled data obtained by sampling second data which is continuous in time.

[0027] In this embodiment, the acquisition unit 11, for example as shown in Figure 1, acquires time-series data (x) included in the learning data D stored in the storage device 2 located outside the learning device 1. t ,y t This time series data is obtained by sampling time-sequentially continuous data (sampled data). t indicates time information (timestamp) at a certain time t, and y t is xt shows the output corresponding thereto.

[0028] In this embodiment, it is described that learning data including time-series data as sampling data is stored in the storage device 2, but it is not limited thereto. Continuous data that is temporally continuous may be stored in the storage device 2. In this case, the acquisition unit 11 may obtain time-series data (x t , y t ) as learning data by sampling the continuous data that is temporally continuous. The storage device 2 is a recording medium capable of storing data, and is configured by, for example, a rewritable non-volatile memory such as a hard disk drive or a solid state drive.

[0029] For example, in the upper and lower rows of (a) in FIG. 2, the process of obtaining time-series data as sampling data from time-series analog data, which is continuous data that is temporally continuous, is conceptually shown. In (a) of FIG. 2, the time-series analog data D, which is the observed time-series data, is divided into {D1, D2,..., D k ,..., D K}, that is, K segments. By sampling the time-series analog data in the segment D k , D c k,m、 D t k,m is conceptually shown to be obtained as learning data. That is, in (a) of FIG. 2, the acquisition unit 11 obtains D c k,m as context set sampling data and D t k,m as target set sampling data as time-series data.

[0030] In other words, D c k,m , which is context set sampling data, is the segment D kIt is obtained by sampling from time-series analog data within a limited range, rather than the entire range. In the example shown in Figure 2(a), segment D k The range is D L k,m , D C k,m , D R k,m It is divided into three ranges. Here, D L k,m ={{x i' ,y i'}∈D k,m |x i'≦ a}, D C k,m ={{x i' ,y i'}∈D k,m |a <x i' <b}、D R k,m ={{x i' ,y i'}∈D k,m |b ≦ x i' It can be expressed as}, where a and b are thresholds. And the divided range D C k,m By sampling from time-series analog data in the context, the context set sampled data D c k,m It has been shown that this can be obtained. On the other hand, the target set sampling data D t k,m This is segment D k It is obtained by sampling from time-series analog data across the entire range.

[0031] Furthermore, in Figure 2(a), the context set sampling data D c k,m The encoder 121 of the NP model 12, described later, is used to generate feature representations. The target set sampling data D t k,m This is used to verify the predicted values ​​output by the decoder 122 of the NP model 12, which will be described later.

[0032] Furthermore, for example, in the upper and lower sections of Figure 3(a), examples of time series data as sampled data are conceptually shown as original data. Note that the original data shown in the upper section of Figure 3(a) is, for example, an example of the first time series data, and the original data shown in the lower section of Figure 3(a) is, for example, an example of the second time series data.

[0033] [1.2 NP Model 12] The NP model 12 that the learning device 1 trains is a neural process model, which is a deep learning model that outputs prediction results that take uncertainty into account using a stochastic process. In this embodiment, the NP model 12 is a neural process model that uses the structure shown in, for example, Convolutional Conditional Neural Processes (ConvCNP) disclosed in Non-Patent Literature 2. Neural Processes (NP) are deep learning models that can predict output values ​​for new inputs conditioned by observed data. In other words, Neural Processes (NP) are deep learning models that can predict the distribution of a function conditioned by observed data.

[0034] Furthermore, the NP model 12 undergoes training of a predictive distribution using a neural process and feature representation learning using self-supervised controlled learning, as described later, by the learning processing unit 13. After training, the NP model 12 can generate multiple predicted values ​​from the same data point in the time series data.

[0035] In this embodiment, the NP model 12 includes an encoder 121 and a decoder 122, as shown in Figure 1.

[0036] [1.2.1 Encoder 121] The encoder 121 is used to extract a time-series representation as a feature representation of the latent space from the time-series data input to the acquisition unit 11. The encoder 121 is a neural network comprising at least a CNN (Convolutional Neural Networks) layer.

[0037] In the upper parts of (b) and (c) of FIG. 2, two random sample points (x c k,m , y c (1,1) ), (x c (1,1) , y c (t,1) ), from the context set sampling data D are taken by the encoder 121, and an example of a process for extracting the feature representation r c (t,1) r c (1,1)、 r c (t,1) of the latent space is shown. Note that ψ shown in (b) of FIG. 2 θ represents the encoder 121 having the model parameter θ. The feature representation r c (1,1)、 r c (t,1) of the latent space extracted by the encoder 121 is aggregated into the feature representation R c 1 of the latent space, as shown in the upper part of (c) of FIG. 2. The feature representation R c 1 of the latent space is the feature representation of the function f c (1,1) , y c (1,1) ), (x c (t,1) , y c (t,1) ) that can represent the input-output relationship of the two sample points.

[0038] Similarly, in the lower parts of (b) and (c) of FIG. 2, two random sample points (x c k,m ), (x c (1,m) , y c (1, m) ), (x c (t, m) , y c (t, m) ) are taken from the context set sampling data D by the encoder 121, and an example of a process for extracting the feature representation r c (1,m)、 r c (t,m) ​​An example of the process by which the latent space feature representation r is extracted by encoder 121 is shown. c (1,m)、 r c (t,m) As shown in the upper part of Figure 2(c), the latent space feature representation R c m It is summarized in the latent space feature representation R. c m This is two sample points (x c (1,m) ,y c (1,m) ), (x c (t,m) ,y c (t,m) A function f that can represent the input-output relationship of ) k This is a characteristic representation of the latent space.

[0039] Note that the processes shown above and below Figure 3(b) and (c) are the same as those shown above and below Figure 2(b) and (c), so the explanation is omitted. Also, in Figure 2(b) and (c), the target set sampling data D t k,m The processing for this is not shown in the diagram, but the context set sampling data D c k,m Since the process is the same as before, the explanation will be omitted.

[0040] [1.2.2 Decoder 122] Decoder 122 outputs predicted values ​​from the feature representations of the latent space extracted and aggregated by encoder 121. More specifically, decoder 122 predicts the time series data distribution that shows the statistical properties of the time series data input to encoder 121 from the feature representations of the latent space, such as embedding vectors extracted and aggregated by encoder 121. Decoder 122 is composed of a different type of neural network than encoder 121.

[0041] In the example shown in the upper part of Figure 2(d), decoder 122 is context This is set sampled data. D ck,m Input X obtained from t Feature representation of the latent space related to 1 R t 1 and input X t From 1, the predicted value Y t 1. Predicted distribution Y t Output the mean and standard deviation of 1. Similarly, in the example shown in the lower part of Figure 2(d), decoder 122 outputs the target set sampling data D t k,m Input X obtained from t m R, a feature representation of the latent space related to this. t m And, input X t m Therefore, the predicted value Y t m That is, the predicted distribution Y t m Output the mean and standard deviation.

[0042] Note that the processes shown above and below Figure 3(d) are the same as those shown in the upper and lower sections of Figure 2(d), so the explanation is omitted.

[0043] [1.3 Learning Processing Unit 13] The learning processing unit 13 includes, for example, a computer including memory and a processor (microprocessor), and realizes the function of performing learning processing by having the processor execute a control program stored in memory. The learning processing unit 13 performs a first learning process 131 on the NP model 12, which learns the predictive distribution using a neural process, and a second learning process 132, which performs feature representation learning using self-supervised controlled learning.

[0044] Here, we will outline feature representation learning using self-supervised contrastive learning and predictive distributions using neural processes.

[0045] [1.3.1 Self-supervised contrastive learning] Figure 4A is a diagram for conceptually explaining self-supervised contrastive learning. Figure 4B is a diagram for conceptually explaining an example of boundary generation by performing the self-supervised contrastive learning shown in Figure 4A. Figure 4A(a) shows labeled cat image 60 and dog image 61. The upper part of Figure 4A(b) shows the transformed images 60a and 60b obtained by transforming cat image 60 with transformation 70. The lower part of Figure 4A(b) shows the transformed image 61a obtained by transforming dog image 61 with transformation 70. Transformation 70 performs data augmentation as a transformation process, and in the example shown in Figure 4A, it is shown as a cropping process. Therefore, images 60a and 60b correspond to a part of cat image 60 and include a part of the cat. Similarly, image 61a corresponds to a part of dog image 61 and includes a part of the dog.

[0046] The upper and lower sections of Figure 4A(c) conceptually illustrate the process of performing self-supervised controlled learning using the transformed images 60a, 60b, and 61a.

[0047] Here, comparative learning is a method that learns so that images with the same original image have similar embedding vectors in the embedding space, while images with different original images have different embedding vectors.

[0048] Self-supervised contrastive learning is a type of self-supervised learning. Self-supervised contrastive learning uses labeled data to encourage embedding vectors of data labeled to the same class to move closer together, and embedding vectors of data labeled to different classes to move further apart.

[0049] More specifically, in the example shown in Figure 4A, images 60a and 60b are cropped images of the same cat image 60, which is labeled to the same class. Therefore, the encoder 71 shown in the upper part of Figure 4A(c) learns to bring the embedding vector 60c, which is the feature representation of image 60a, and the embedding vector 60d, which is the feature representation of image 60b, closer together. On the other hand, images 60b and 61a are cropped images of a cat image 60 and a dog image 61, which are labeled to different classes. Therefore, the encoder 71 shown in Figure 4A(c) learns to move the embedding vector 60d, which is the feature representation of image 60b, and the embedding vector 61b, which is the feature representation of image 61a, further apart.

[0050] Conceptually, the self-supervised controlled learning shown in Figure 4A corresponds to learning to move images 61a and 60b further apart and images 60b and 60a closer together, as shown in Figure 4B, which makes it easier to draw a boundary line to distinguish between dogs and cats.

[0051] By performing self-supervised contrastive learning in this way, it is possible to generate a model that can recognize differences in appearance from a small amount of image data. In other words, by performing self-supervised contrastive learning, the model can learn data consistency, so it is possible to obtain a highly accurate model with high recognition performance even from a small amount of labeled data.

[0052] [1.3.2 Predictive distributions using neural processes] Next, we will explain predictive distributions using neural processes.

[0053] Figure 5A is a diagram that conceptually explains the prediction distribution by a neural process. Figure 5A shows the structure of the neural process model and an example of its processing. Figure 5B shows an example of the prediction result by the neural process shown in Figure 5A.

[0054] As shown in Figure 5A, the structure of the neural process model 82, like the variational autoencoder, consists of two types of neural networks: an encoder e and a decoder d. This configuration makes the neural process model 82 a deep learning model that can perform regression on time-series data while considering uncertainty, similar to a Gaussian process, which is one of the machine learning methods other than deep learning.

[0055] In the neural process model 82, as shown in Figure 5A, the encoder e has j input / output pairs of observation points {(x1,y1), ..., (x j ,y j Enter )} for each observation point (x j ,y j ) may be different data points in the time series data. In the neural process model 82, each observation point (x j ,y j The output of encoder e for ) r j After the calculation, the calculated output r is sent to aggregator a. j The data is aggregated to calculate r, which is the latent representation vector of multiple observation points. In the example shown in Figure 5A, the output r is sent to the aggregator a. j The average r is calculated as the latent representation vector of multiple observation points. Then, the decoder d receives the average r, which is the latent representation vector of multiple observation points, as the predicted point x. T By inputting this along with the neural process model 82, the output y considering the observation point is generated. T This makes it possible to predict the output y. T The prediction is the mean μ of the distribution (regression) predicted from multiple observation points. yT and standard deviation σ yT This is done by outputting the following.

[0056] Figure 5B shows the predicted distribution (regression) from multiple observation points indicated by "×" as prediction results of the neural process model 82 shown in Figure 5A, with the region labeled "Confidence" being displayed. .figureThe solid line shown in 5B represents the average of the predicted distribution (regression) from multiple observation points indicated by "×". The "Confidence" region indicates the uncertainty of the predicted distribution (regression) from multiple observation points. Furthermore, the predicted distribution (regression) from multiple observation points can be said to represent the statistical properties of those multiple observation points.

[0057] By using such a neural process model 82, it is possible to output prediction results that take uncertainty into account using a stochastic process.

[0058] [1.3.3 Details of the first learning process 131 and the second learning process 132] The learning processing unit 13 performs a first learning process 131 on the NP model 12, which learns a predictive distribution using a neural process, and a second learning process 132 on the NP model 12, which learns feature representations using supervised comparative learning.

[0059] More specifically, the learning processing unit 13 performs a first learning process 131 to train the NP model 12 so that it predicts a first time series data distribution that shows the statistical properties of the first time series data and a second time series data distribution that shows the statistical properties of the second time series data, based on the first and second time series data. The learning processing unit 13 also performs a second learning process 132 to perform feature representation learning using supervised comparative learning with a comparative learning algorithm. Specifically, as the second learning process 132, the learning processing unit 13 performs learning to bring together multiple first sampled data points generated by sampling from the first time series data distribution as positive samples. The learning processing unit 13 also performs learning to bring together multiple second sampled data points generated by sampling from the second time series data distribution as positive samples. On the other hand, the learning processing unit 13 performs a second learning process 132, which involves learning to separate negative samples from a plurality of first sample data points and a plurality of second sample data points. In this embodiment, the learning processing unit 13 performs the first learning process 131 and the second learning process 132 by batch learning. The learning processing unit 13 may also perform the first learning process and the second learning process simultaneously.

[0060] Here, using Figure 6, the learning method of this disclosure, i.e., the first learning process 131 and the second learning process 132 according to this embodiment, will be conceptually explained.

[0061] Figure 6 is a diagram for conceptually explaining the first learning process 131 and the second learning process 132 according to this embodiment.

[0062] The upper and lower sections of Figure 6(a) show two different sets of original data. The original data shown in the upper and lower sections of Figure 6(a) corresponds to the original data shown in Figure 3(a). In other words, the original data shown in the upper section of Figure 6(a) is, for example, an example of the first time series data, and the original data shown in the lower section of Figure 6(a) is, for example, an example of the second time series data.

[0063] The upper and lower sections of Figure 6(b) conceptually show the distribution (predictive data distribution) that represents the predicted statistical properties of the original data, based on the first training process performed on the NP model 12. The dotted line shown in the predictive distribution in Figure 6(b) is, for example, the mean of the predictive distribution. Therefore, in the first training process of the NP model 12, the parameters of the NP model 12 are learned so that it can output a distribution (predictive data distribution) that represents the predicted statistical properties of the original data, based on the original data.

[0064] The upper and lower sections of Figure 6(d) conceptually illustrate that self-supervised comparative learning is performed using data generated by sampling from the predictive distribution learned by the first learning process. In other words, in this embodiment, instead of conventional data augmentation processes such as trimming, the NP model 12 learned by the first learning process is used to predict the predictive distribution of time series data. Then, data generated by sampling from the region representing the predicted predictive distribution is used for self-supervised comparative learning. This makes it possible to handle time series data in self-supervised comparative learning.

[0065] More specifically, the data generated by sampling the predicted distribution predicted by the neural process from the original data in the upper part of Figure 6(a) have the same source data. Therefore, a second learning process is performed to perform comparative learning so that the embedding vectors of these data are closer together. Similarly, the data generated by sampling the predicted distribution predicted by the neural process from the original data in the lower part of Figure 6(a) have the same source data. Therefore, a second learning process is performed to perform comparative learning so that the embedding vectors of these data are closer together.

[0066] On the other hand, the data generated by sampling the predicted distribution predicted by the neural process from the original data in the upper part of Figure 6(a) and the data generated by sampling the predicted distribution predicted by the neural process from the original data in the lower part of Figure 6(a) are based on different source data. Therefore, a second learning process is performed to differentiate the embedding vectors of the data generated by sampling the predicted distribution predicted by the neural process from the original data in the upper and lower parts of Figure 6(a).

[0067] In the examples shown in Figures 2 and 3, the latent space feature representation R shown in Figures 2 and 3(c) is used. c m , R c 1 corresponds to the embedding vector of the data generated by sampling the predicted distribution predicted by the neural process shown in Figure 6(c). Also, the predicted value Y shown in Figures 2 and 3(d) t 1, Y t m That is, the predicted distribution Y t 1, Y t m The mean and standard deviation of the data correspond to the mean and standard deviation of the predictive distribution.

[0068] Furthermore, as shown in Figure 2(c), the latent space feature representation R c m , R c 1 is learned using the first error function for comparative learning, denoted by Lc(θ). In other words, in the second learning process, the first error function is used to obtain the latent space feature representation R extracted by the encoder 121. c m , R c The NP model 12 is trained to move closer together when the original data is the same, and further apart when the original data is different.

[0069] Distribution Y shown in Figure 2(d) t 1, Y t m is -log p(Y k,m |ψ θ (Rk,m )(X k,m The log-likelihood function used in neural process learning, as shown by ), is used as the second error function for learning. In other words, in the first learning process, the output of decoder 122 and distribution Y are used by the second error function. t 1, Y t m The theoretical value is compared with the encoder 121, and the latent space feature representation R with less error is obtained. c m , R c The NP model 12 is trained so that 1 is extracted.

[0070] More specifically, when performing the first and second learning processes, the learning processing unit 13 uses an error function that modifies the second error function by adding a term of the first error function used in the contrast learning algorithm to the term of the second error function. The first error function is an error function that reduces the error in the case of positive samples and increases the error in the case of negative samples, and is also called the contrast error function. The second error function is an error function related to the error of the prediction result used in the neural process model. In this way, the learning processing unit 13 performs the first and second learning processes using the final error function that combines the first and second error functions.

[0071] Furthermore, Lc(θ), shown in Figure 2(c), that is, the first error function, can be expressed by (Equation 1).

[0072]

number

[0073] Here, O(k,n) represents the sampled data generated by sampling from the time series data distribution.

number

[0074]

number

[0075] Also,

number

[0076] Furthermore, in (Equation 1), sim represents cosine similarity, and generally

number

[0077] Note that (Equation 1) represents the normalized Temperature CrossEntropy (NT-Xent) error function used in SimCLR, but applied to data generated by sampling instead of data generated by data augmentation.

[0078] The final error function, which combines the first and second error functions used by the learning processing unit 13, can be expressed by the following (Equation 2).

[0079]

number

[0080] As shown in (Equation 2), the final error function obtained by combining the first and second error functions used by the learning processing unit 13 is an error function obtained by adding a term obtained by multiplying the first error function by the hyperparameter h to the second error function. The first term of the error function shown in (Equation 2) can also be called the regression term, and the second term can be called the contrast term. By using the error function shown in (Equation 2), the learning processing unit 13 can simultaneously perform the first learning process 131, which learns the predictive distribution using a neural process, and the second learning process 132, which performs feature representation learning using supervised contrast learning.

[0081] [2.1 Processing Procedure for Learning Device 1] Next, we will explain the processing procedure of the learning device 1 using Figure 7.

[0082] Figure 7 shows pseudocode of algorithm 1, which is the processing procedure of the learning device 1 according to this embodiment. Algorithm 1 shown in Figure 7 corresponds to the processing of the learning device 1 shown in Figure 3 and is performed, for example, by the processor of the learning device 1. Note that the functions such as f([:a]) and f shown in Figure 7 represent the original data (or original time-series analog data) shown in Figure 3, and indicate that f([:a]) and f are different original data.

[0083] Figure 7(i) specifies that two data points, oc1 and oc2, are obtained by sampling from the original time-series analog data or original data f([:a]). It also specifies that two data points, ot1 and ot2, are obtained by sampling from the original time-series analog data or original data f. Note that the process in Figure 7(i) corresponds to, for example, the process in which two sample points (data) are obtained by sampling from the original data shown in the upper and lower sections of Figure 3(a), as shown in the upper and lower sections of Figure 3(b).

[0084] Figure 7(ii) specifies that the process involves extracting and aggregating the latent space feature representations rc1 and rc2 from the two data points oc1 and oc2 obtained in the process of Figure 7(i) using an encoder. It also specifies that the process involves deriving the latent space feature representations rt1 and rt2 from the two data points ot1 and ot2 obtained in the process of Figure 7(i) and the latent space feature representations rc1 and rc2. Note that the process in Figure 7(ii) corresponds to the process shown in the upper and lower sections of Figure 3(b), for example.

[0085] Figure 7(iii) specifies that the predicted distribution p_y is obtained using a decoder from rt derived in the process of Figure 7(ii) and the input data xt. In other words, the process of Figure 7(iii) corresponds to the process of outputting the predicted distribution shown in the upper and lower sections of Figure 3(d), for example.

[0086] Figure 7(iv) specifies that the learning process is performed using backpropagation with an error function that includes regression and control terms. In other words, the process in Figure 7(iv) corresponds to, for example, the second learning process 132 which comparatively learns the feature representations of the latent space shown in the upper and lower sections of Figure 3(c), and the process which reduces the error by comparing the predicted distribution shown in the upper and lower sections of Figure 3(d) with the theoretical value.

[0087] By executing algorithm 1 as defined in this way, the learning device 1 can simultaneously perform a first learning process 131 that learns a predictive distribution using a neural process, and a second learning process 132 that performs feature representation learning using supervised comparative learning.

[0088] [2.2 Operation of Learning Device 1] Next, we will explain the operation of the learning device 1 configured as described above.

[0089] Figure 8 is a flowchart illustrating the overview of the operation of the learning device 1 according to this embodiment.

[0090] First, the learning device 1 acquires training data that includes the first time series data and the second time series data which is different from the first time series data (S101).

[0091] Next, the learning device 1 performs a first learning process (S102) to train the NP model 12 to predict the first time series data distribution, which represents the statistical properties of the first time series data, and the second time series data distribution, which represents the statistical properties of the second time series data, from the first time series data and the second time series data.

[0092] Next, the learning device 1 performs a second learning process (S103) that uses a contrast learning algorithm to perform feature representation learning using supervised contrast learning. Specifically, the learning processing unit 13 performs a second learning process 132 that brings together multiple first sample data generated by sampling from the first time series data distribution as positive samples. The learning processing unit 13 also performs a second learning process 132 that brings together multiple second sample data generated by sampling from the second time series data distribution as positive samples. On the other hand, the learning processing unit 13 also performs a second learning process 132 that brings together multiple first sample data and multiple second sample data as negative samples.

[0093] [3 Effects, etc.] As described above, the learning device 1 and learning method according to this embodiment make it possible to realize a learning method that can handle time-series data in self-supervised learning by combining the framework of self-supervised controlled learning with the learning framework of a neural process model.

[0094] More specifically, according to the learning device 1 and learning method of this embodiment, by using a neural process instead of data augmentation processing, a neural process model can be trained (first learning process) to generate different data from the same source data in time series data. Furthermore, the neural process model can also be trained (second learning process) by performing comparative learning on the feature representations in the latent space of two data generated by the neural process. Therefore, according to the learning device 1 and learning method of this embodiment, by performing the first and second learning processes, it is possible to learn time series data that takes uncertainty into account within the framework of a neural process, and to learn data consistency using the framework of self-supervised comparative learning. This not only enables the realization of a learning method that can handle time series data in self-supervised learning, but also enables training to create a highly accurate model from a small amount of time series data and a small amount of labels for that data.

[0095] Furthermore, according to the learning device 1 and learning method of this embodiment, the first learning process and the second learning process are performed using an error function that combines a first error function used in the comparative learning algorithm and a second error function related to the error of the prediction result used in the neural process model. This makes it possible to simultaneously perform the first learning process, which learns the prediction distribution by the neural process, and the second learning process, which performs feature representation learning using self-supervised comparative learning, for the neural process model to be learned.

[0096] (Example of experiment) The effectiveness of the learning method and other aspects disclosed herein was verified using the MIT-BIH Atrial Fibrillation (AFDB), IMS Bearing, and Urban8K datasets, and the verification results are described as an experimental example.

[0097] Figure 9 shows the results of evaluating the performance of the disclosed model using the dataset related to the experimental example. ContrNP(ours) shown in Figure 9 corresponds to the disclosed model, i.e., the NP model 12 described above. Figure 9 also shows the results of evaluating the performance of SimCLR and the performance of supervised learning as comparative examples.

[0098] The AFDB dataset contains 25 electrocardiogram (ECG) data points. Each data point is approximately 10 hours long. The AFDB dataset also includes four classes: atrial fibrillation, atrial flutter, atrioventricular junctional rhythm, and all other rhythms. The reason for choosing the AFDB dataset in this experiment is its long data duration and the fact that its characteristics change over time (the classes alternate).

[0099] Furthermore, the IMS Bearing dataset contains data collected from a run-to-failure experiment in which four bearings rotating at 2000 rpm on a shaft under a 6000 lbs load failed and were then addressed. The IMS Bearing dataset is divided into five classes, each representing the health of the bearings (initial, normal, impending failure, etc.). The reason for choosing the IMS Bearing dataset in this experiment was to evaluate its performance on long, noisy industrial time-series data.

[0100] The Urban8K dataset also includes 8,732 audio files of varying sizes, each under 4 seconds in length. The Urban8K dataset consists of audio files categorized into 10 classes, such as children playing music, car horns, dog barks, and street music.

[0101] The encoder 121 used in this experiment is constructed using a Convolutional Neural Network (CNN). As shown in Figure 9, performance was evaluated by comparing accuracy and average accuracy (AUPRC: area under the precision-recall curve). The accuracy and average accuracy values ​​represent the average after 5, 5, and 10 runs, respectively.

[0102] As shown in Figure 9, for example, on the AFDB dataset, the performance of the ContrNP(ours) model disclosed in this disclosure is more than 10% better than the performance of SimCLR, which is used as a comparative example.

[0103] From the above, it can be seen that the model ContrNP(ours) disclosed in this disclosure, despite having an error function for comparative learning similar to that of the comparative model SimCLR, outperforms SimCLR on all datasets. Furthermore, it can be seen that the performance of the model ContrNP(ours) disclosed in this disclosure, while not reaching the level of performance achieved by supervised learning, is close to that level.

[0104] Figure 10 shows the accuracy when ContrNP(ours) is trained with different label rates for the AFDB dataset related to the experimental example. As a comparison, Figure 10 also shows the accuracy when SimCLR is trained and the accuracy when supervised learning is performed.

[0105] As can be seen from Figure 10, ContrNP(ours), i.e., the model disclosed herein, shows an accuracy of over 80% even when the label rate is as low as a few percent, outperforming SimCLR. Furthermore, when the label rate is around 15% or higher, the disclosed model shows performance close to that of supervised learning.

[0106] (Possibility of other embodiments) Although the learning device and learning method of this disclosure have been described in the embodiments above, the entities and devices on which each process is performed are not particularly limited. Processing may be performed by a processor or the like embedded in a specific device located locally. Alternatively, processing may be performed by a cloud server or the like located in a different location from the local device.

[0107] This disclosure is not limited to the embodiments described above. For example, other embodiments realized by arbitrarily combining the components described herein, or by excluding some of the components, may also be considered embodiments of this disclosure. Furthermore, modifications obtained by applying various modifications to the above embodiments that a person skilled in the art could conceive of, without departing from the spirit of this disclosure, i.e., the meaning of the language used in the claims, are also included in this disclosure.

[0108] Furthermore, this disclosure also includes the following cases:

[0109] (1) Specifically, the above-mentioned device is a computer system consisting of a microprocessor, ROM, RAM, hard disk unit, display unit, keyboard, mouse, etc. A computer program is stored in the RAM or hard disk unit. The microprocessor operates according to the computer program, thereby enabling each device to perform its function. Here, the computer program is composed of a combination of multiple instruction codes that indicate commands to the computer in order to achieve a predetermined function.

[0110] (2) Some or all of the components constituting the above-described device may be made up of a single system LSI (Large Scale Integration). The system LSI is a multi-functional LSI manufactured by integrating multiple components onto a single chip, and specifically, it is a computer system comprising a microprocessor, ROM, RAM, etc. A computer program is stored in the RAM. The system LSI achieves its function by operating the microprocessor in accordance with the computer program.

[0111] (3) Some or all of the components constituting the above-described device may consist of a removable IC card or a standalone module attached to each device. The IC card or module is a computer system consisting of a microprocessor, ROM, RAM, etc. The IC card or module may include the above-described multi-function LSI. The microprocessor operates according to a computer program, thereby enabling the IC card or module to perform its function. The IC card or module may be tamper-resistant.

[0112] (4) The present disclosure may also be the methods described above. Alternatively, it may be a computer program that implements these methods using a computer, or a digital signal consisting of the computer program.

[0113] (5) The disclosure may also be a computer program or a digital signal recorded on a computer-readable recording medium, such as a flexible disk, hard disk, CD-ROM, MO, DVD, DVD-ROM, DVD-RAM, BD (Blu-ray® Disc), semiconductor memory, etc. Alternatively, the digital signal may be recorded on one of these recording media.

[0114] Furthermore, this disclosure may also describe transmitting the computer program or digital signal via telecommunications lines, wireless or wired communication lines, networks such as the Internet, data broadcasting, etc.

[0115] Furthermore, the present disclosure may also provide a computer system comprising a microprocessor and memory, wherein the memory stores the computer program, and the microprocessor operates in accordance with the computer program.

[0116] Furthermore, the program or digital signal may be implemented by another independent computer system by recording and transferring it on the recording medium, or by transferring the program or digital signal via the network or the like. [Industrial applicability]

[0117] This disclosure can be used in learning methods, learning devices, and programs for learning highly accurate models from a small amount of data and a small number of labels, and in particular in learning methods, learning devices, and programs that can learn highly accurate models capable of handling time series data through self-supervised learning. [Explanation of symbols]

[0118] 1. Learning device 2 Storage device 11 Acquisition Department 12 NP Model 13 Learning Processing Unit 121 encoders 122 Decoder 131 First Learning Process 132 Second Learning Process

Claims

1. A learning method using batch learning performed by a computer, Training data is obtained that includes a first time series data and a second time series data different from the first time series data. A first learning process is performed to train a neural process model, which is a deep learning model that outputs prediction results that take uncertainty into account using a stochastic process, so as to predict a first time series data distribution that shows the statistical properties of the first time series data and a second time series data distribution that shows the statistical properties of the second time series data from the first time series data and the second time series data. A second learning process is performed using a contrast learning algorithm, which involves learning to bring together multiple first sample data generated by sampling from the first time series data distribution as positive samples, learning to bring together multiple second sample data generated by sampling from the second time series data distribution as positive samples, and learning to move apart the multiple first sample data and the multiple second sample data as negative samples. The neural process model includes an encoder that outputs a feature representation of the latent space from time-series data. In the second learning process described above, the object of learning using the comparative learning algorithm is the feature representation output from the encoder, The parameters learned in the second learning process are the parameters of the encoder. Learning methods.

2. The aforementioned first time series data is time series sample data obtained by sampling time-sequenced first data, The aforementioned second time series data is time series sample data obtained by sampling the second data which is continuous in time. The learning method according to claim 1.

3. The first learning process and the second learning process are performed simultaneously. When performing the first learning process and the second learning process, The error function used is one in which the second error function is modified by adding a term of the first error function used in the contrast learning algorithm, which reduces the error in the case of positive samples and increases the error in the case of negative samples, to the term of the second error function related to the error of the prediction result used in the neural process model. The learning method according to claim 1 or 2.

4. A learning device for performing learning by batch learning, An acquisition unit that acquires training data including a first time series data and a second time series data different from the first time series data, A first learning process is performed to train a neural process model, which is a deep learning model that outputs prediction results that take uncertainty into account using a stochastic process, so as to predict a first time series data distribution that shows the statistical properties of the first time series data and a second time series data distribution that shows the statistical properties of the second time series data from the first time series data and the second time series data. The system includes a learning processing unit that performs a second learning process using a contrast learning algorithm, which involves learning to bring together multiple first sampled data points generated by sampling from the first time series data distribution as positive samples, learning to bring together multiple second sampled data points generated by sampling from the second time series data distribution as positive samples, and learning to move apart the multiple first sampled data points and the multiple second sampled data points as negative samples. The neural process model includes an encoder that outputs a feature representation of the latent space from time-series data. In the second learning process described above, the object of learning using the comparative learning algorithm is the feature representation output from the encoder, The parameters learned in the second learning process are the parameters of the encoder. Learning device.

5. A program that causes a computer to execute a learning method using batch learning, Training data is obtained that includes a first time series data and a second time series data different from the first time series data. A first learning process is performed to train a neural process model, which is a deep learning model that outputs prediction results that take uncertainty into account using a stochastic process, so as to predict a first time series data distribution that shows the statistical properties of the first time series data and a second time series data distribution that shows the statistical properties of the second time series data from the first time series data and the second time series data. Using a contrast learning algorithm, the computer is instructed to perform a second learning process in which it learns to bring together multiple first sample data generated by sampling from the first time series data distribution as positive samples, brings together multiple second sample data generated by sampling from the second time series data distribution as positive samples, and moves apart the multiple first sample data and the multiple second sample data as negative samples. The neural process model includes an encoder that outputs a feature representation of the latent space from time-series data. In the second learning process described above, the object of learning using the comparative learning algorithm is the feature representation output from the encoder, The parameters learned in the second learning process are the parameters of the encoder. program.

Images