Learning device, identification device, learning method, and program

Abstract

This learning device comprises an input feature embedding unit, a clustering unit, and a model learning unit. The input feature embedding unit performs embedding processing for embedding, in a low-dimensional vector representation, an input feature, which is optional input. The clustering unit performs clustering processing for classifying, by similar data source features, data source features each indicating an input feature data source, and outputs, as result of the clustering processing, a cluster, which is a collection of similar data source features. For each cluster output by the clustering unit, the model learning unit learns a model so that vector representations corresponding to data source features which belong to the same cluster are closer to each other.

Description

Learning device, identification device, learning method, and program

[0001] The present invention relates to a learning device, a classifier, a learning method, and a program, and more particularly to a learning device, a classifier, a learning method, and a program that enable zero-shot learning even when the number of instances per dataset is small.

[0002] In recent years, machine learning of classification models has been conducted for multiple datasets with differences in data distribution, such as individual differences. One example is emotion recognition from human biometric signals. In such cases, it is necessary to model each individual's sensibilities, so models are trained individually. However, in reality, there is a limit to the amount of data that can be obtained from individuals, so there is a high demand for methods that can learn the relationship between data and classes even when no data is available for some classes.

[0003] As such a learning method, there is a learning category called zero-shot learning (ZSL), which uses auxiliary information about classes to indirectly learn the relationship between classes and data for which no data has been obtained, as described in Non-Patent Document 1.

[0004] Andrea Frome, Greg S. Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, Tomas Mikolov. DeViSE: A Deep Visual-Semantic Embedding Model. Advances in Neural Information Processing Systems (NIPS2013), Vol. 26, 2013

[0005] However, the above-mentioned zero-shot learning is a method that deals with missing classes, and there is a problem in that it cannot learn a model when the number of instances of classes that are not missing is also small.

[0006] The present invention has been made in light of these circumstances, and aims to provide a learning device, a classification device, a learning method, and a program that, given a set of small datasets with missing classes and information describing each dataset, clusters the datasets in advance and introduces contrastive learning using the clustering results as labels, thereby alleviating both the problems of insufficient data and missing classes, and thereby enabling zero-shot learning even when the number of instances per dataset is small.

[0007] A first aspect of the present invention is a learning device that includes an input feature embedding unit that performs an embedding process to embed input features, which are arbitrary inputs, into a low-dimensional vector representation; a clustering unit that performs a clustering process to classify data source features, which indicate the data source of the input features, into groups of similar data source features, and outputs clusters, which are sets of similar data source features, as a result of the clustering process; and a model learning unit that learns a model for each cluster output by the clustering unit so that vector representations corresponding to data source features belonging to the same cluster are similar.

[0008] A second aspect of the present invention is an identification device comprising: an input feature embedding unit that performs an embedding process to embed input features, which are arbitrary inputs, into a low-dimensional vector representation; a clustering unit that performs a clustering process to classify data source features, which indicate the data source of the input features, into groups of similar data source features, and outputs clusters, which are sets of similar data source features, as a result of the clustering process; a model selection unit that selects an appropriate model for the cluster output by the clustering unit from trained models; and an identification unit that identifies the vector representation obtained by the embedding process by the input feature embedding unit, using the model selected by the model selection unit.

[0009] A third aspect of the present invention is a learning method implemented by a learning device, in which a processor of the learning device performs an embedding process to embed input features, which are arbitrary inputs, into a low-dimensional vector representation; a clustering process to classify data source features, which indicate the data sources of the input features, into groups of similar data source features, and output clusters, which are sets of similar data source features, as a result of the clustering process; and a model learning process to train, for each output cluster, the vector representations corresponding to data source features belonging to the same cluster so that they are similar.

[0010] A fourth aspect of the present invention is a program for causing a computer to function as each unit included in the learning device of the first aspect.

[0011] According to the present invention, given a set of small datasets with missing classes and information representing each dataset, the datasets are clustered in advance and contrastive learning is introduced using the clustering results as labels, thereby alleviating both the problems of insufficient data and missing classes. This makes it possible to provide a learning device, a classification device, a learning method, and a program that enable zero-shot learning even when the number of instances per dataset is small.

[0012] Fig. 1 is a functional block diagram showing an example of the configuration of a learning and identification device to which a learning and identification method according to an embodiment of the present invention is applied. Fig. 2 is a simplified diagram showing the hardware configuration of a general computer. Fig. 3 is a data flow showing the flow of processing by the learning device of the learning and identification device according to an embodiment of the present invention. Fig. 4 is a data flow showing the flow of processing by the identification device of the learning and identification device according to an embodiment of the present invention.

[0013] Hereinafter, embodiments of the present invention will be described with reference to the drawings. The drawings are schematic or conceptual, and the relationship between the thickness and width of each part, the size ratio between parts, etc., are not necessarily the same as those in reality. Furthermore, even when the same part is shown, the dimensions and ratios may be different depending on the drawing. In this specification and each drawing, elements similar to those described above with reference to the previous drawings are designated by the same reference numerals, and detailed and redundant explanations will be omitted as appropriate.

[0014] FIG. 1 is a functional block diagram showing an example of the configuration of a learning and identification device to which a learning and identification method according to an embodiment of the present invention is applied.

[0015] The learning and classification device 10 includes a learning device 20, a classification device 30, and a model database unit 40.

[0016] The learning and classification device 10 can be realized by a computer such as a PC.

[0017] FIG. 2 is a simplified diagram showing the hardware configuration of a typical computer.

[0018] Computer 100 receives some kind of input from the outside, processes it, and outputs the results to the outside. Input device 120 is responsible for input, and output device 140 is responsible for output. CPU (Central Processing Unit) 110 controls the overall flow of data and processing, and performs calculations and other processes. In Figure 2, solid arrows represent the flow of data, and dashed arrows represent the flow of control.

[0019] A program describing the processing procedure is required for the computer 100 to perform processing. The program contains a series of instructions and data used by the instructions, all of which are temporarily stored in the memory 130.

[0020] The CPU 110 retrieves instructions from the memory 130 one by one, interprets the instructions, and operates in accordance with the instructions.

[0021] Therefore, the learning and classification device 10 according to this embodiment is realized by the operation of the CPU 110 in accordance with the programs stored in the memory 130 together with the learning device 20 and the classification device 30 .

[0022] The computer 100 may also include a storage device (not shown), in which the model database unit 40 may be stored.

[0023] Before returning to FIG. 1 and describing the details of the learning device 20 and the identification device 30, the definitions of the terms "input feature," "data source," and "data source feature" used in this specification will first be explained.

[0024] The learning and classification device 10 according to this embodiment does not depend on the modality of the input data. Therefore, in this specification, the term "input feature" refers to data that belongs to any input feature space, such as an image, a sound, sensor data, table data, or a graph.

[0025] In this specification, the term "data source" refers to the source of data. For example, in the case of emotion recognition of multiple people using biosignals as input, each individual whose biosignal is measured is considered to be a different data source.

[0026] In this specification, the term "data source characteristics" refers to information that expresses the characteristics of a data source, such as demographic information such as an individual's date of birth, or the results of responses to a questionnaire.

[0027] (Learning Device) As illustrated in FIG. 1, the learning device 20 includes a learning input unit 21, a clustering unit 22, an input feature embedding unit 23, and a model learning unit 24.

[0028] FIG. 3 is a data flow showing the flow of processing by the learning device of the learning and classification device according to the embodiment of the present invention.

[0029] The learning input unit 21 receives input features, which are arbitrary inputs, and data source features from multiple data sources (not shown) (S1). The data source features are information indicating from which data source the input features are obtained. The learning input unit 21 then processes the set of input features and the data source features as inputs, and generates a set D of input features and labels. ｗ and the data source features B.

[0030] The labels of the input features are assumed to be manually given. The set of input features and labels is given from W different data sources and expressed as follows: Here, N ｗ is the number of instances in the wth dataset.

[0031] x ｗｉ is an input feature that belongs to the input feature space X. Also, the label y ｗｉ is a set of classes Y that differs for each data source. ｗ and satisfies the following:

[0032] Here, K is the number of classes. That is, for each data set D ｗ Although it does not necessarily cover all instances of the K classes, it indicates that all classes are covered when all data sets are combined. Furthermore, the features of each data source are summarized as follows:

[0033] In this way, the learning input unit 21 generates a set D of input features and labels. ｗ and the data source features B.

[0034] The clustering unit 22 performs a clustering process to classify the data source features B output by the learning input unit 21 into similar data source features, and outputs clusters consisting of similar data source features (e.g., first cluster, second cluster, third cluster, etc.) as a result of the clustering process C (S2).

[0035] This allows the clustering unit 22 to cluster the data source features B and clarify data sources that have similar tendencies.

[0036] Any known clustering technique such as the k-means method can be used for the clustering technique. The cluster to which the data source w obtained by clustering belongs is called c. ｗ Then, the clustering result can be expressed as follows: where: where M indicates the number of clusters that can be determined arbitrarily so that M<W. The clustering unit 22 outputs the clustering result c in this way.

[0037] The input feature embedding unit 23 embeds the input features and label set D output by the learning input unit 21. ｗ The input feature is extracted from the input feature space X by The process of embedding the vectors in a dimensional space, i.e., vector representation, is performed, and a set E of embedded vectors and labels is output (S3).

[0038] To perform this process, the input feature embedding unit 23 performs the model This model can use any model, such as a matrix that performs linear transformation or a neural network that performs linear or nonlinear transformation.

[0039] The input feature embedding unit 23 embeds the input feature x into the model φ. ｗｉ By giving the embedding vector e ｗｉ = φ(x ｗｉ ) can be obtained. The input feature embedding unit 23 repeats this operation for all instances in all data sources to obtain a set of embedding vectors and labels. Output.

[0040] The model learning unit 24 learns a model for each cluster output by the clustering unit 22 so that the vector representations corresponding to the data source features (e.g., A, B, C, ...) belonging to the same cluster (e.g., the first cluster) are close to each other for each class label, and outputs a model (e.g., parameters) for each cluster (e.g., the first cluster, the second cluster, the third cluster, ...) as a result of the learning.

[0041] Specifically, the model learning unit 24 receives as input the set E of embedding vectors and labels output from the input feature embedding unit 23 and the clustering result c of the data source output from the clustering unit 22, introduces contrastive learning into the existing zero-shot learning, and performs a process of learning a model that can predict the labels of the embedding vectors, thereby generating a set of trained discriminative models. To perform this process, the model learning unit 24 outputs the same number of discrimination models as the number of clusters M. Here, the model parameters of the model φ are ０ and the model The model parameters are θ ｍ and the set is At this time, the model learning unit 24 calculates the objective function The model parameters are estimated by optimizing the following two rules: indicates the expected value. （ｍ） teeth Among them, the clustering result c of data source w ｗ But, c ｗ = m ｗ It is a set of. Any hyperparameter can be set as a coefficient that determines the weight of each term in this objective function. Two rules are explained below.

[0042] (Rule 1) The embedding model and the discriminative model are trained by zero-shot learning. This is because the loss function L of the well-known zero-shot learning technique is ＺＳＬ This is achieved by minimizing (·).

[0043] (Rule 2) The embedding model is trained by contrastive learning using class labels and clustering results of data sources as labels. This is based on the assumption that the embedding representations of instances with the same label in data sources that are determined to be similar by clustering are similar. Therefore, for a pair of class labels and clustering results (y ｗｉ , c ｗ ) are positive samples, and the others are negative samples. ＣＬ Minimize (・).

[0044] Any optimization method such as the gradient method, the stochastic gradient method, Adam, etc. can be applied to optimize the objective function. When the gradient method is used, the parameters are repeatedly updated according to the following formula in the sth optimization step (S4). where γｓ is the learning rate parameter. The gradient of the objective function may be calculated using a calculated function or may be calculated numerically.

[0045] As explained using Figure 2, the operation of each component of the learning device 20 can be constructed as a program, which can be installed and executed on the computer 100 used as the learning device 20, but it can also be distributed via a network.

[0046] The model database unit 40 stores the set of trained discriminant models output from the model training unit 24 (S5). is available to the identification device 30 .

[0047] 1, the classification device 30 includes an input unit 31, a clustering unit 32, an input feature embedding unit 33, a model selection unit 34, and a classification unit 35. The clustering unit 32 and the input feature embedding unit 33 have the same functions as the clustering unit 22 and the input feature embedding unit 23. Furthermore, the clustering unit 32 and the input feature embedding unit 33 can be shared by the clustering unit 22 and the input feature embedding unit 23, so that they can be formed into the same block.

[0048] FIG. 4 is a data flow showing the flow of processing by the classification device of the learning classification device according to the embodiment of the present invention.

[0049] Similarly to the learning input unit 21, the input unit 31 receives input features and data source features, which are arbitrary inputs, from a plurality of data sources (not shown) (S11). The input unit 31 then performs processing using the set of input features and the data source features as inputs, and generates a set D of input features and labels. ＊ , and data source feature b ＊ The input feature labels are assumed to be manually given. The set of input features and labels is It is expressed as N ＊ is the number of instances in the dataset. is an input feature that belongs to the input feature space X. Also, the label corresponding to the input feature is the class set Y ＊ It belongs to Also, the data source features are ＊ It is expressed as:

[0050] In this way, the input unit 31 receives the set D of input features and labels. ＊ , and data source feature b ＊ Output.

[0051] The clustering unit 32 can be the clustering unit 22 that has already been trained by the learning device 20. Therefore, the clustering unit 32 performs a clustering process on the data source features output from the input unit 31 to classify the data source features into groups of similar data source features, as described for the clustering unit 22, and outputs clusters (e.g., a first cluster, a second cluster, a third cluster, ...) consisting of similar data source features as a result of the clustering process (S12).

[0052] As a result, the clustering unit 32 extracts the data source feature b ＊ is used as input, and the data source feature b ＊ The data source is clustered from the data source, and the clustering result c ＊ can be output.

[0053] The input feature embedding unit 33 can be the input feature embedding unit 23 that has already been trained by the training device 20, and the set D of input features and labels output by the input unit 31 ＊ is used as input, and the input features are extracted from the input feature space X to a lower-dimensional The embedding process is performed in a dimensional space, i.e., a vector representation, and a set of embedding vectors and labels E ＊ (S13) The details of the process are the same as those of the input feature embedding unit 23, so a duplicate explanation will be avoided.

[0054] The model selection unit 34 refers to the model database unit 40 and selects the clustering result c output by the clustering unit 32. ＊Pre-trained model corresponding to is selected and output (S14).

[0055] The identification unit 35 identifies the vector representation output from the input feature embedding unit 33 using the model selected by the model selection unit 34. Specifically, the identification unit 35 identifies the set E of embedded vectors and labels output from the input feature embedding unit 33 using the model selected by the model selection unit 34. ＊ and the trained discrimination model output from the model selection unit 34. and the trained discriminant model The embedding vector is given to , and the process of outputting a set of predicted categorical distributions is performed. For example, the discrimination unit 35 outputs the trained discrimination model A set of embedding vectors and labels Given an instance of , the predicted categorical distribution The classification unit 35 repeats this operation for all instances to obtain a set of predicted categorical distributions. Output.

[0056] As described above, in the identification device 30, the clustering unit 32 and the input feature embedding unit 33 can be the clustering unit 22 and the input feature embedding unit 23 that have already been trained by the learning device 20. In the learning device 20, the clustering unit 22 and the input feature embedding unit 23 are used in a state where the model parameters are not fixed, but in the identification device 30, they are used in a state where they have already been trained by the learning device 20 and the model parameters are fixed. Therefore, the identification unit 35 can perform identification with higher accuracy.

[0057] As described above, according to a learning and classification device to which a learning and classification method according to an embodiment of the present invention is applied, given a set of small datasets with missing classes and information representing each dataset, the datasets are pre-clustered, and contrastive learning using the clustering results as labels is implemented, thereby alleviating both the problems of insufficient data and missing classes. This makes zero-shot learning possible even when the number of instances per dataset is small. In other words, zero-shot learning of a model is possible even when some classes are missing and the number of instances of non-missing classes is also small.

[0058] The present invention is not limited to the above-described embodiments, and can be embodied by modifying the components within the scope of the gist of the present invention. Furthermore, various inventions can be created by appropriately combining multiple components disclosed in the above-described embodiments. For example, some components may be omitted from all the components shown in the embodiments. Furthermore, components from different embodiments may be appropriately combined.

[0059] REFERENCE SIGNS LIST 10 Learning and classification device 20 Learning device 21 Learning input unit 22 Clustering unit 23 Input feature embedding unit 24 Model learning unit 30 Classification device 31 Input unit 32 Clustering unit 33 Input feature embedding unit 34 Model selection unit 35 Classification unit 40 Model database unit 100 Computer 120 Input device 130 Memory 140 Output device

Claims

1. A learning device comprising: an input feature embedding unit that performs an embedding process to embed input features, which are arbitrary inputs, into a low-dimensional vector representation; a clustering unit that performs a clustering process to classify data source features that indicate the data source of the input features into similar data source features, and outputs clusters that are sets of similar data source features as a result of the clustering process; and a model learning unit that learns a model for each cluster output by the clustering unit so that the vector representations corresponding to the data source features belonging to the same cluster become closer.

2. An identification device comprising: an input feature embedding unit that performs an embedding process to embed input features, which are arbitrary inputs, into a low-dimensional vector representation; a clustering unit that performs a clustering process to classify data source features, which indicate the data source of the input features, into groups of similar data source features, and outputs clusters, which are sets of similar data source features, as a result of the clustering process; a model selection unit that selects an appropriate model for the cluster output by the clustering unit from trained models; and an identification unit that uses the model selected by the model selection unit to identify the vector representation obtained by the embedding process by the input feature embedding unit.

3. A learning method implemented by a learning device, in which a processor of the learning device performs the following steps: performing an embedding process to embed input features, which are arbitrary inputs, into a low-dimensional vector representation; performing a clustering process to classify data source features, which indicate the data source of the input features, into similar data source features, and outputting clusters, which are sets of similar data source features, as a result of the clustering process; and learning a model for each output cluster so that the vector representations corresponding to the data source features belonging to the same cluster are similar.

4. A program for causing a computer to function as each of the units provided in the learning device according to claim 1.

Images