JP7440798B2 - Learning device, prediction device, learning method and program - Google Patents

Description

The present invention relates to a learning device, a prediction device, a learning method, and a program technique.

A framework called supervised learning is generally used for predictive model learning using machine learning. Supervised learning is a framework that prepares a large number of pairs of certain data and correct class labels for that data, and learns the relationships between the data and class label pairs.

In order to implement supervised learning, it is necessary to prepare pairs of large amounts of data and class labels, but creating these pairs is basically expensive. Therefore, a method is sometimes taken in which a model learned in an area where supervised data already exists (hereinafter referred to as a "domain") is utilized in a target domain. For example, when recognizing handwritten characters, after training a discriminator using digital font data for which supervised data is relatively easy to obtain, the discriminator is trained using handwritten character data with little (or no supervised data). Retraining methods may be used.

However, the original domain for which learning was performed (hereinafter referred to as the "source domain"; in the case of the previous example, the digital font data) and the target domain (hereinafter referred to as the "target domain"; in the case of the previous example, the handwritten character data) The data generation distribution may differ between the two. FIG. 6 is a diagram schematically showing such a problem. In FIG. 6, the region surrounded by solid lines is the source domain 10, the region surrounded by broken lines is the target domain 20, and the straight line is the identification boundary 30. For example, even if the character "a" is the same, the shape of the digital font and the handwritten character may be significantly different. When the generation distributions are different, the discrimination boundary 30 learned in the source domain 10 as shown in FIG. 6 may be unreliable in the target domain 20. In such a case, a problem arises in that the learned model cannot achieve the expected classification accuracy in the target domain 20. In this way, a learning problem in which there are differences between domains is called a domain adaptation problem.

Conventionally, in order to solve such domain adaptation problems, the following known techniques exist. In the technique disclosed in Patent Document 1, the process is performed from the source domain to the target domain in such a way as to minimize the value of MMD, which is the perceived distribution distance between the sample generation distribution in the source domain and the sample generation distribution in the target domain. A conversion rule is learned. Then, data in the original domain is transformed using the learned transformation rule, and model learning is performed by supervised learning using the transformed data in the original domain.

Non-Patent Document 1 describes a feature extractor that projects source domain data and target domain data onto a feature space that makes it difficult to identify the domain, and a feature extractor that projects source domain data and target domain data into a feature space that makes domain identification difficult, and The relationship with the assigned class label is learned at the same time. Making it difficult to distinguish between the data of the source domain and the data of the target domain in the feature space means making the generation distributions of both closer to each other in the feature space. Such processing may mean, for example, changing the state of FIG. 6 to the state of FIG. 7. This improves the prediction accuracy for the target domain data of the model obtained by performing supervised learning on the source domain data.

In Non-Patent Document 2, the common feature space learned in Non-Patent Document 1 is learned using a feature extractor and two discriminators connected to the feature extractor. It is known that the model learned in Non-Patent Document 2 has higher prediction accuracy for target domain data than the model learned in Non-Patent Document 1.

Various collateral problems may arise depending on the differences between the source and target domains. One of the accompanying problems is the problem that occurs when data other than the class given to the source domain exists in the target domain. Taking the case of handwritten character recognition mentioned above as an example, even though the digital font data only contains "a", "i", and "u", the handwritten character data contains "e" and "o". Such a problem arises when it is included. Classes that are labeled by the original domain are called known classes (in the previous example, "A", "I", and "U"), and other classes are called unknown classes (in the previous example, "A", "I", and "U"). ``E'', ``O''). Normally, a classifier that has undergone supervised learning predicts that even if data belonging to an unknown class is input, the data belongs to one of the known classes. Such an operation may cause a problem in that the accuracy of character recognition decreases.

In addition, there are other problems as follows. Normally, it is assumed that the source domain and the target domain each consist of a single domain. However, both the source domain and the target domain may be formed by multiple domains. For example, when handwritten character data is written by multiple different individuals or using different writing instruments, there is a possibility that the source domain and the target domain are formed by multiple domains. In this case, since the generation distribution changes, it can be considered that a plurality of domains exist within the target domain. When a domain is formed by a plurality of domains, a problem arises in which the method described in Non-Patent Document 1 cannot achieve the expected prediction accuracy.

Non-patent document 4 is a document related to a technique for dealing with the problem of multiple domains inherent in the original domain. In the technique disclosed in Non-Patent Document 4, features that make it difficult to distinguish between each domain inherent in the source domain and the target domain are learned. On the other hand, there is Non-Patent Document 5 as a document related to a technique for dealing with the problem that a target domain includes multiple domains. In the technique disclosed in Non-Patent Document 5, features that make it difficult to identify a domain among multiple domain regions within a target domain are learned.

JP 2019-101789 Publication

Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In David Blei and Francis Bach, editors, Proceedings of the 32nd International Conference on Machine Learning (ICML15), pages 1180-1189. JMLR Workshop and Conference Proceedings, 2015. 2 Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, Tatsuya Harada. Maximum Classifier Discrepancy for Unsupervised Domain Adaptation. The 31th IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2018), 2018 K. Saito, S. Yamamoto, Y. Ushiku, and T. Harada. Open set domain adaptation by backpropagation. In The European Conference on Computer Vision (ECCV), September 2018. 2 Han Zhao, Shanghang Zhang, Guanhang Wu, Jose MF ´ Moura, Joao P Costeira, and Geoffrey J Gordon. Adversarial multiple source domain adaptation. In Advances in Neural Information Processing Systems, pages 8568-8579, 2018. 1, 2, 5 Z. Chen, J. Zhuang, X. Liang, and L. Lin. Blending-target domain adaptation by adversarial meta-adaptation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2248-2257, 2019.

In order to solve the domain adaptation problem in which the generation distribution is different between the source domain and the target domain, and the various incidental problems that occur along with it, we have proposed methods such as those in Non-patent Document 3, Non-patent Document 4, and Non-patent Document 5, respectively. techniques have been proposed. However, although each technique shows good performance for the concomitant problem it considers, it is not effective for other concomitant problems.

For example, even if the technique of Non-Patent Document 4, which deals with a problem where multiple domains are inherent in the source domain, is applied to a problem where multiple domains are inherent in the target domain, sufficient performance cannot be obtained. do not have.

Generally, it is rare that it is possible to know in advance what kind of problems exist in the target of processing. Therefore, it is difficult to judge what kind of problem the technology should be applied to. Furthermore, in cases where multiple of the above-mentioned problems coexist, there is also the problem that it becomes difficult to apply the technology.

In view of the above circumstances, the present invention has been made in view of such problems, and aims to provide a technique that achieves good performance for a wider range of domain-related problems.

One aspect of the present invention includes: a feature extractor that outputs feature amounts of input data; and a plurality of discriminators that obtain probability of belonging to a known class and an unknown class for the data based on the feature amounts. an unknown class discriminator that determines whether the data is an unknown class based on the attribution probability obtained by the discriminator; and an unknown class discriminator that determines whether the data is an unknown class based on the attribution probability obtained by the discriminator; The feature extractor calculates the identification discrepancy value by using the identification discrepancy evaluator that outputs the identification discrepancy value that indicates the difference in probability, and the data that is not the unknown class and to which no teacher label is attached. a learning unit that iteratively learns the parameters of the feature extractor and the plurality of classifiers so as to decrease the value and increase the value of the discrimination inconsistency degree for the plurality of classifiers; It is a device.

One aspect of the present invention provides a feature extractor that outputs a feature amount of input data based on a parameter obtained by the above learning device, and a feature extractor that outputs a feature amount of input data based on a parameter obtained by the above learning device; and a discriminator that obtains the probability of belonging of the data to a known class and an unknown class based on the data.

One aspect of the present invention includes a feature extraction step of outputting a feature amount of input data using a feature extractor, and a feature extraction step of outputting a feature amount of input data using a plurality of discriminators to identify a known class for the data based on the feature amount. an identification step of acquiring respective belonging probabilities to unknown classes; an unknown class identification step of determining whether the data is an unknown class based on the acquired belonging probabilities; an identification discrepancy evaluation step of outputting a value of the classification discrepancy degree indicating the difference in the respective attribution probabilities obtained by the classifiers; Repetition of the parameters of the feature extractor and the plurality of classifiers so that the value of the degree of discrimination inconsistency is decreased for the feature extractor, and the value of the degree of discrimination inconsistency is increased for the plurality of classifiers. This learning method includes a learning step for performing learning.

One aspect of the present invention is a program for operating a computer as the learning device described above.

The present invention has been made in view of such problems, and makes it possible to achieve good performance for a wider range of domain-related problems.

FIG. 1 is a diagram schematically showing the present embodiment. FIG. 1 is a diagram schematically showing the present embodiment. FIG. 1 is a functional block diagram showing an example of a learning device 100 according to the present embodiment. It is a functional block diagram showing an example of a prediction device 200 according to the present embodiment. 3 is a flowchart illustrating an example of the operation of the learning device 100. FIG. 2 is a diagram showing an example of conventional technology. FIG. 2 is a diagram showing an example of conventional technology.

<Summary>
First, the outline of this embodiment will be explained. This embodiment operates appropriately even when there is a problem in which an unknown class exists (hereinafter referred to as the "first problem"). Furthermore, this embodiment solves the problem that the data in each domain is only partially labeled (hereinafter referred to as the "second problem") and the problem that the domain attribution information of the data is unknown (hereinafter referred to as the "second problem"). The configuration may be configured to operate appropriately even if the third problem exists. Furthermore, the system may be configured to operate appropriately even when a plurality of problems among these three incidental problems are present.

More specifically, it is as follows. FIGS. 1 and 2 are diagrams schematically showing the present embodiment. In FIGS. 1 and 2, the region surrounded by solid lines is the source domain 10, the region surrounded by broken lines is the target domain 20, and the straight line is the identification boundary 30. A line segment 40 indicates information identified as a boundary between a known class and an unknown class. Arrow 50 indicates configuring domain adaptation.

This embodiment deals with the first problem of the existence of an unknown class by identifying and specifying data belonging to the unknown class from among data to which no teacher label has been given. In this embodiment, for the second and third problems, labeled data is considered as the source domain, and data that is not given a teacher label and belongs to a known class is considered as the target domain. Deal with it by configuring adaptations.

<Example of configuration of learning device>
Next, the configuration of the learning device according to this embodiment will be explained. FIG. 3 is a functional block diagram showing an example of the learning device 100 according to this embodiment. The learning device 100 is configured using, for example, an information processing device such as a personal computer or a server device. The learning device 100 includes a control section 90, an unknown class information storage section 130, and a learning result storage section 140. The control unit 90 is configured using a processor such as a CPU (Central Processing Unit) and a memory. The control unit 90 controls the feature extractor 101, the first classifier 102, the second classifier 103, the classification loss evaluation unit 104, the unknown class classifier 105, the classification discrepancy evaluation unit 106, and the learning by the processor executing the program. 107. Note that all or part of each function of the control unit 90 may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). The above program may be recorded on a computer-readable recording medium. Computer-readable recording media include portable media such as flexible disks, magneto-optical disks, ROMs, CD-ROMs, semiconductor storage devices (for example, SSDs: Solid State Drives), and hard disks and semiconductor storages built into computer systems. It is a storage device such as a device. The above program may be transmitted via a telecommunications line.

The learning device 100 operates by acquiring data from the supervised data storage section 110 and the unsupervised data storage section 120. The supervised data storage unit 110 is configured using a device or medium capable of storing data, such as a storage device such as a magnetic hard disk device or a semiconductor storage device, or a recording medium such as a CD-ROM. The supervised data storage unit 110 stores a supervised data set. A supervised data set is a data set that is given a desired class label. The unsupervised data storage unit 120 is configured using a device or medium capable of storing data, such as a storage device such as a magnetic hard disk device or a semiconductor storage device, or a recording medium such as a CD-ROM. The unsupervised data storage unit 120 stores unsupervised data sets. An unsupervised data set is a data set to which a desired class label has not been assigned.

The feature extractor 101 receives a supervised data set and an unsupervised data set as input, and extracts a feature vector from each data. The feature extractor 101 outputs the extracted feature vectors to the first classifier 102 and the second classifier 103. The feature extractor 101 operates based on a function having parameters that can extract such feature vectors. A feature vector is, for example, a vector representing the features of data. In other words, the feature vector represents the features of necessary data as a vector having n-dimensional elements. n may be any integer value, for example n=512. Note that although the feature vector will be described as having a vector format for convenience, the format is irrelevant to the main point of the present invention and can take any format. Each time the feature extractor 101 outputs a feature vector, it reads the parameters stored in the learning result storage unit 140 and outputs the feature vector.

The first classifier 102 receives the feature vector output by the feature extractor 101 as input. The first classifier 102 outputs an estimated value of the probability of belonging to each class and the unknown class (hereinafter referred to as "estimated belonging probability") for the original data of the input feature vector. Estimated belonging probability is a probability representing the likelihood that data belongs to each known class and unknown class. The first classifier 102 operates based on a function having parameters that can output such an estimated belonging probability. Each time the first discriminator 102 outputs the estimated belonging probability, it reads the parameters stored in the learning result storage unit 140 and outputs the estimated belonging probability.

The second classifier 103 receives the feature vector output by the feature extractor 101 as input. The second classifier 103 outputs an estimated value of the probability of belonging to each class and the unknown class (estimated belonging probability) for the original data of the input feature vector. The second classifier 103 operates based on a function having parameters that can output such an estimated belonging probability. Every time the second discriminator 103 outputs the estimated belonging probability, it reads the parameters stored in the learning result storage unit 140 and outputs the estimated belonging probability. Note that the same feature vector is input to the first classifier 102 and the second classifier 103.

Any functions can be used as the functions applied to the feature extractor 101, the first classifier 102, and the second classifier 103 as long as they are differentiable with respect to the parameters. In this embodiment, a CNN (Convolutional Neural Network) is used. However, CNN is only one example, and there is no need to be limited to this.

The identification loss evaluation unit 104 includes data to be processed, information indicating whether or not the data to be processed is an unknown class, and a first discriminator 102 and a second discriminator 103 for the data to be processed. The output estimated attribution probability and the desired attribution probability for the data to be processed (hereinafter referred to as "teacher attribution probability") are received as input. The discrimination loss evaluation unit 104 obtains a value of a discrimination loss function (hereinafter referred to as "discrimination loss evaluation value") which is a first loss function representing these differences. The teacher attribution probability is the attribution probability according to the class label that is the correct answer during learning.

The unknown class classifier 105 receives as input the data to be processed and the estimated belonging probabilities output by the first classifier 102 and the second classifier 103 for the data to be processed. The unknown class discriminator 105 identifies whether the data to be processed is an unknown class. The unknown class discriminator 105 records information indicating the identification result (hereinafter referred to as "unknown class information") in the unknown class information storage section 130. The information recorded in the unknown class information storage section 130 is used by the identification loss evaluation section 104 and the identification mismatch evaluation section 106.

The identification discrepancy evaluation unit 106 receives as input the data to be processed and the estimated attribution probabilities output by the first classifier 102 and the second classifier 103 for the data to be processed. The identification discrepancy evaluation unit 106 obtains a value indicating the degree of discrepancy between the estimated attribution probabilities of the first classifier 102 and the second classifier 103 (hereinafter referred to as "identification discrepancy evaluation value").

The learning unit 107 receives as input the classification loss function obtained by the classification loss evaluation unit 104 and the classification mismatch degree evaluation value obtained by the classification discrepancy evaluation unit 106. The learning unit 107 performs iterative learning of the parameters of the feature extractor 101, first classifier 102, and second classifier 103 using the input values. The learning unit 107 records the parameters of the feature extractor 101, first classifier 102, and second classifier 103 obtained through iterative learning in the learning result storage unit 140. Iterative learning regarding the feature extractor 101 is performed so that both the classification loss evaluation value and the classification mismatch evaluation value become small. Iterative learning regarding the first classifier 102 and the second classifier 103 is performed so that the classification loss evaluation value becomes small and the classification mismatch degree evaluation value becomes large.

<Example of configuration of prediction device>
Next, the configuration of the prediction device according to this embodiment will be explained. FIG. 4 is a functional block diagram showing an example of the prediction device 200 according to this embodiment. The prediction device 200 is configured using, for example, an information processing device such as a personal computer or a server device. The prediction device 200 includes a control section 91 and a storage section 230. The control unit 91 is configured using a processor such as a CPU and a memory. The control unit 91 functions as a feature extractor 201 and a discriminator 202 when a processor executes a program. Note that all or part of each function of the control unit 91 may be realized using hardware such as an ASIC, a PLD, or an FPGA. The above program may be recorded on a computer-readable recording medium. Computer-readable recording media include, for example, portable media such as flexible disks, magneto-optical disks, ROMs, CD-ROMs, semiconductor storage devices (such as SSDs), and storage devices such as hard disks and semiconductor storage devices built into computer systems. It is a device. The above program may be transmitted via a telecommunications line.

The storage unit 230 is configured using a storage device such as a magnetic hard disk device or a semiconductor storage device. The storage unit 230 stores parameters as learning results obtained through iterative learning performed by the learning unit 107 of the learning device 100.

When the feature extractor 201 receives data to be processed (data to be predicted) 240, it reads parameters from the storage unit 230 and operates based on the parameters. The feature extractor 201 outputs a feature vector for the data 240 to be processed. The discriminator 202 reads parameters from the storage unit 230 and operates based on the parameters. The classifier 202 calculates an estimated belonging probability for the data 240 to be processed, based on the feature vector obtained by the feature extractor 201. The output of the classifier 202 may be the estimated belonging probability itself for each class regarding the data 240 to be processed, or may be information indicating the prediction result of which class the data 240 belongs to.

<Example of operation of learning device>
FIG. 5 is a flowchart showing an example of the operation of the learning device 100. Next, an example of the operation of the learning device 100 will be described. The learning device 100 receives the supervised data set 110 and the unsupervised data set 120, and executes the learning processing routine shown in FIG.

First, the control unit 90 of the learning device 100 reads one or more supervised data sets 110 and unsupervised data sets 120 (step S101). Next, the control unit 90 makes a branch determination as to whether or not the number of repetitions of learning is equal to or less than a predetermined number of times (step S102). If the number of repetitions is less than or equal to the scheduled number of times, the process of step S103 is executed. On the other hand, if the number of repetitions is greater than the scheduled number of times, the process of step S104 is executed.

Here, the significance of the branch processing in step S102 will be explained. This branching process changes the method of identifying unknown classes. The first classifier 102 and the second classifier 103 learn to identify (K+1) classes, which is a total of K known classes and unknown classes. However, for known classes, a set of data and its teacher attribution probability is obtained as supervised data, whereas for unknown classes, it is unclear which data is the unknown class. Therefore, if the number of repetitions is less than the planned number, unknown classes are identified for the unsupervised data, and the results are recorded as an identification history. On the other hand, if the number of repetitions is greater than the planned number, the first classifier 102 and the second classifier 103 are trained to identify (K+1) classes, and the results of identifying unknown classes are recorded as a classification history.

If the number of iterations is less than the predetermined rank, the teacher belonging probability of the unknown class can be estimated by identifying the unknown class, but this may include errors. Therefore, by learning the classification of (K+1) classes while recording the classification history, it is possible to identify unknown classes with fewer errors.

In step S103, the feature extractor 101, the first classifier 102, the second classifier 103, and the unknown class classifier 105 are applied to the supervised data set 110 and the unsupervised data set 120, and the classification loss evaluation value, An identification inconsistency evaluation value and a determination as to whether the class is an unknown class or not are obtained.

In step S104, the unknown class identification history for the unsupervised data set 120 is read. Then, in step S105, the feature extractor 101, the first classifier 102, the second classifier 103, and the unknown class classifier 105 are applied to the supervised data set 110, the unsupervised data set 120, and the unknown class identification history. As a result, an identification loss evaluation value, an identification inconsistency evaluation value, and a determination as to whether or not the class is an unknown class are obtained.

When the process of step S103 or step S105 is finished, the learning unit 107 determines the values of the parameters of the feature extractor 101, the first classifier 102, and the second classifier 103 based on the classification loss evaluation value and the classification inconsistency evaluation value. (values recorded in the learning result storage unit 140) are updated (step S106).

The parameters of the feature extractor 101, first classifier 102, and second classifier 103 are stored in the learning result storage unit 140. Next, the unknown class discriminator 105 records the identification result as to whether the data is unknown class data obtained in step S103 or S105 in the unknown class information storage unit 130 (step S107).

Then, the control unit 90 determines whether the termination condition is satisfied (step S108). If the termination condition is satisfied (step S108-YES), the control unit 90 terminates the process. If the termination condition is not satisfied (step S108-NO), the control unit 90 returns to step S101 and repeats the process.

Through the iterative learning described above, the parameters of the feature extractor 101, first classifier 102, and second classifier 103 are learned. Regarding the feature extractor 101, learning is performed using the identification loss evaluation value and the identification inconsistency evaluation value so that the identification loss evaluation value and the identification inconsistency evaluation value become small. The identification loss function outputs a smaller value as the degree of similarity between the estimated attribution probability of the data output by the first classifier 102 and the second classifier 103 and the given teacher attribution probability of the data is higher. The identification discrepancy evaluation value indicates the difference between the classifiers in the estimated attribution probability of the data output by the first classifier 102 and the second classifier 103. Further, regarding the first classifier 101 and the second classifier 102, learning is performed so that the classification loss evaluation value is small and the classification mismatch degree evaluation value is large.

[Details of each process]
Next, details of processing of each processing unit of the learning device 100 will be explained.
[If the number of repetitions is less than the planned number]
In step 102, each process of the classification loss evaluation section 104, the unknown class classifier 105, and the classification mismatch evaluation section 106 when the number of repetitions is less than or equal to the planned number of repetitions will be explained.

The discrimination loss function calculates the degree of similarity between the estimated attribution probability of the data output by the first classifier 102 and the second classifier 103 using the feature vector output from the feature extractor 101 as input, and the given teacher attribution probability of the data. The higher the value, the smaller the value output. The discrimination loss function corresponds to Equation 2 and Equation 3, which will be described later. Further, the value corresponds to the identification loss evaluation value.

[Processing of identification loss evaluation unit]
The feature extractor 101 is realized by using a function F that receives data x as input, outputs a feature vector f, and has a parameter φ. The first classifier 102 can be expressed as a function having a parameter θ1 that receives the feature vector f as an input and outputs an estimated belonging probability y1. The second classifier 103 can be expressed as a function having a parameter θ2 that receives the feature vector f as an input and outputs an estimated belonging probability y2. A function that implements the first classifier 102 and second classifier 103 can be expressed as a probability function as shown in Equation 1 below using a function F that implements the feature extractor 101. Note that i is used as a subscript to distinguish between two classifiers.

The formula is the probability that yi will appear given φ, θi, and x. Desirable feature extractor 101, first classifier 102, and second classifier 103 are such that when data s from a supervised data set is given, supervised attribution probability t to each class appears. That is, the feature extractor 101, the first classifier 102, and the second classifier 103 are used to determine the probability of belonging to which a correct class can be identified. When the appearance probability of the teacher attribution probability t corresponding to the data s is p(s, t), it is sufficient for learning to determine the parameters φ and θi so that the following equation 2 becomes small.

Eb[a] is the expected value of probability b of a. In the case of this embodiment, since supervised data is obtained from a supervised data set, the expected value is approximately replaced in the form of a summation as shown in Equation 3 below.

Note that S and T are each a set of one or more data and corresponding teacher attribution probabilities. Equation 3 is the discrimination loss function in an example of this embodiment, and the value obtained by evaluating this for arbitrary S and T is the discrimination loss evaluation value.

By reducing Equation 3 with respect to φ, θ1, and θ2, it is possible to obtain a desirable feature extractor 101, first classifier 102, and second classifier 103 that can output t for s. There are various methods to obtain such φ, θ1, and θ2. Simply put, if the function F that realizes the feature extractor and the probability functions representing the first classifier 102 and the second classifier 103 are differentiable with respect to the respective parameters φ, θ1, θ2, then the local It is known that it can be minimized. Therefore, in an example of the present embodiment, the feature extractor 101 is a function that outputs a feature vector f of the data when data x is input, is differentiable with respect to φ, and is used as a first discriminator. 102 and the second discriminator 103 may be selected from functions that satisfy the conditions that the function outputs the estimated membership probabilities y1 and y2 by inputting the feature vector f, and that it is differentiable with respect to θ1 and θ2, respectively. .

[Processing of identification discrepancy evaluation unit]
The discrimination inconsistency evaluation value for certain estimated belonging probabilities p1 and p2 is expressed as shown in the following equation 4, when p1k and p2k represent the belonging probabilities of estimated belonging probabilities p1 and p2 for class k, respectively. Here, K represents the number of known classes to be identified, and K+1 represents an unknown class that does not fall under any of the known classes.

The identification discrepancy evaluation unit 106 evaluates the degree of discrepancy between the estimated belonging probabilities y1 and y2 output by the first classifier 102 and the second classifier 103 for the data u of the unsupervised data set 120. That is, the identification discrepancy evaluation unit 106 uses the identification discrepancy degree evaluation value of the estimated attribution probability in equation 4 to calculate the first discriminator for the appearance probability p(u) of data u in the unsupervised data set shown in equation 5 below. 102 and the second classifier 103 are output.

Eb[a] is the expected value of probability b of a. In the case of this embodiment, since the unsupervised data is obtained from an unsupervised data set, the expected value is approximately replaced in the form of a sum as shown in Equation 6 below.

Note that U is one or more pieces of data. Equation 6 is the identification mismatch degree in an example of this embodiment, and the value obtained by evaluating this for any U is the identification mismatch degree evaluation value.

[Unknown class classifier processing]
Estimated attribution probabilities y1 and y2 output by the first classifier 102 and the second classifier 103 for data x can be expressed using Equation 1 above. The information entropy H (y | It is expressed as follows.

Determination of whether the unsupervised data u of the unsupervised data set is unknown class data is determined by whether the value of information entropy shown in Equation 4 is larger than a predetermined threshold value σ. That is, the identification y _u,e of whether or not the unsupervised data u at the e-th iteration is unknown class data is expressed as shown in Equation 8 below.

[If the number of repetitions is greater than a certain value]
The process of dividing the unsupervised data set into the known class data set U _I and the unknown class data set U _O in step S105 will be described. The identification result as to whether or not data u of the unsupervised data set is unknown class data at the number of iterations t is stored in the unknown class information storage unit 130 as y _u,t in step S107, which will be described later. . In step S104, the past T identification results are read from the unknown class information storage unit 130, and for the unsupervised data set data u that has been identified as unknown class data more than T/2 times in the past, the unknown class data set U Data belonging to _O , and other data belonging to known class data set U _I. That is, when the unsupervised data set is U at the number of iterations e, U is divided into a known class data set U _I and an unknown class data set U _O according to Equations 9 and 10 below.

Next, the evaluation process related to step S105 will be explained. The processing of the identification loss evaluation section 104 and the identification mismatch evaluation section 106 is substantially the same as step S103, which is the processing when the number of repetitions is less than a certain value.

[Processing of identification loss evaluation unit]
The discrimination loss evaluation unit 104 obtains a discrimination loss evaluation value by calculating the sum of the union of the supervised data and its teacher attribution probability set (S, T) and the unknown class data set U _O. That is, the evaluation value of the identification loss evaluation unit 104 is expressed in the form of Equation 11 below.

[Processing of identification discrepancy evaluation unit]
Regarding the process of the identification discrepancy evaluation unit 106, the identification discrepancy degree evaluation value is obtained by performing the same processing as the evaluation process of Equation 6 of the identification discrepancy evaluation unit 106 in step _S103 on the data of the known class data set U I. seek. That is, the identification inconsistency degree evaluation value output by the identification inconsistency evaluation unit 106 in step S105 is obtained by equation 12 below.

[Processing of unknown class classifier]
Estimated attribution probabilities y ₁ and y ₂ output by the first classifier 102 and the second classifier 103 for data x can be expressed using Equation 1 above. The average estimated attribution probability y can be determined from the estimated attribution probabilities y ₁ and y ₂ output by the first classifier 102 and the second classifier 103. To determine whether unsupervised data u of an unsupervised data set is unknown class data, for the average estimated attribution probability y, among the attribution probabilities for each identified class, the attribution probability for the K+1 class, which is an unknown class, is the highest. If the data is high, it is determined that it is unknown class data, and if not, it is determined that it is not unknown class data. That is, the identification y _u,e of whether the unsupervised data u at the e-th iteration is unknown class data is expressed as shown in Equation 13 below.

[Learning process]
The learning process of the learning unit 107 in step S106 will be explained. Regarding the feature extractor 101, learning processing is performed so that the values of the discrimination loss evaluation value L _s and the discrimination inconsistency degree evaluation value L _adv become small. Regarding the first classifier 102 and the second classifier 103, learning processing is performed so that the classification loss evaluation value L _s is small and the classification mismatch degree evaluation value is large. Specifically, the problems shown in Equations 14, 15, and 16 are sequentially optimized.

Here, the functions of the feature extractor 101, the first classifier 102, and the second classifier 103 are set so that the classification loss evaluation value Ls and the classification discrepancy evaluation value L _adv are differentiable with respect to the parameters θ ₁ , θ ₂ , and φ. Since we chose , it is possible to learn using the error gradient effect method.

The expected effects of the above learning will be explained. First, regarding L _s , minimizing the parameters θ ₁ , θ ₂ , and φ produces the effect of improving recognition accuracy based on supervised data, similar to general discrimination learning.

Regarding L _adv , learning is performed so that the values of the parameters θ ₁ and θ ₂ are increased, and the parameter φ is minimized. Details regarding the effect of this learning are as described in Non-Patent Document 4. The distribution of supervised data and the distribution of unsupervised data in the space of features output by the feature extractor 101 become close to each other. As the distribution in the feature space becomes closer, it becomes possible for a classifier trained on supervised data to perform highly accurate recognition when recognizing unsupervised data.

However, if the distribution of supervised data and the distribution of unsupervised data are brought closer together in the feature space simply by learning similar to Non-Patent Document 4, the unknown class data among the unsupervised data will also be brought closer together. At this time, the unsupervised unknown class data approaches the supervised data, and is identified as one of the classes of known class data that is originally inappropriate. In this embodiment, unknown class data is detected, and in step S105, data detected as unknown class data is not used for L _adv evaluation. This prevents the above-mentioned inappropriate supervised data distribution from becoming close to the unknown class data distribution, and makes it possible to learn to detect unknown class data as unknown class data.

[Parameter storage processing]
After parameter learning, the parameters θ ₁ , θ ₂ , and φ are stored in the learning result storage unit 140 in the process related to step S107.

The process of saving the identification result as to whether the unsupervised data is unknown class data in the process of step S108 will be described. As for the identification history of whether the data u of the unsupervised data set at the number of iterations e is unknown class data, when the number of iterations e is less than a certain value, y _u,e is obtained by the process of step S103. In addition, the identification history of whether the data u of the unsupervised data set at the number of iterations e is unknown class data is that if the number of iterations e is greater than a certain value, y _u,e is obtained by the process of step S305. . In step S108, the identification results y _u,e are stored in the unknown class information storage unit 130 for each data u of the unsupervised data set.

The learning process from steps S101 to S108 described above may be repeated until the termination condition is satisfied.

Any information may be used as the termination condition. For example, "until it is repeated a predetermined number of times", "until the value of the objective function no longer changes beyond a certain level", "until the accuracy of evaluation data prepared separately from the training data no longer changes beyond a certain level", etc. good.

(Modified example)
Either or both of the supervised data storage section 110 and the unsupervised data storage section 120 may be included in the learning device 100. Either or both of the unknown class information storage section 130 and the learning result storage section 140 may be provided outside the learning device 100. If it is provided externally, data may be acquired through communication such as TCP/IP.

The learning device 100 may be implemented using one information processing device, or may be distributed and implemented among multiple information processing devices.

Although the embodiments of the present invention have been described above in detail with reference to the drawings, the specific configuration is not limited to these embodiments, and includes designs within the scope of the gist of the present invention.

The present invention is applicable to learning devices.

100...Learning device, 101...Feature extractor, 102...First classifier, 103...Second classifier, 104...Discrimination loss evaluation section, 105...Unknown class classifier, 106...Identification discrepancy evaluation section, 107...Learning section , 200...prediction device

Claims (7)

a feature extractor that outputs feature quantities of input data;
a plurality of classifiers that obtain probability of belonging to a known class and an unknown class for the data based on the feature amount;
an unknown class classifier that determines whether the data is an unknown class based on the belonging probability obtained by the classifier;
an identification inconsistency evaluation unit that outputs an identification inconsistency degree value indicating a difference in each of the attribution probabilities obtained by the plurality of classifiers for the data;
Using data that is not the unknown class and to which no teacher label has been assigned, the value of the identification discrepancy is reduced for the feature extractor, and the identification discrepancy for the plurality of classifiers is a learning unit that repeatedly learns parameters of the feature extractor and the plurality of classifiers so as to increase the value;
Equipped with
Further comprising a discriminant loss evaluation unit that outputs, for the data, a value of a discriminant loss function that indicates a smaller value as the degree of similarity between the attribution probability and a given teacher attribution probability of the data is higher;
The learning unit performs the discrimination using the feature extractor and the plurality of classifiers using data to which a teacher label is assigned and data that is an unknown class and does not have a teacher label. The learning device further performs iterative learning of the parameters so as to reduce the value of the loss function . a feature extractor that outputs feature quantities of input data;
a plurality of classifiers that obtain probability of belonging to a known class and an unknown class for the data based on the feature amount;
an unknown class classifier that determines whether the data is an unknown class based on the belonging probability obtained by the classifier;
an identification inconsistency evaluation unit that outputs an identification inconsistency degree value indicating a difference in each of the attribution probabilities obtained by the plurality of classifiers for the data;
Using data that is not the unknown class and to which no teacher label has been assigned, the value of the identification discrepancy is reduced for the feature extractor, and the identification discrepancy for the plurality of classifiers is a learning unit that repeatedly learns parameters of the feature extractor and the plurality of classifiers so as to increase the value;
Equipped with
The unknown class classifier is a learning device that makes a determination based on past determination results when the number of times of iterative learning in the learning section is greater than a predetermined number of times . 3. The learning device according to claim 1, wherein the unknown class classifier makes a determination based on the belonging probability when the number of times of iterative learning in the learning section is less than or equal to a predetermined number. A feature extractor that outputs feature amounts of input data based on parameters obtained by the learning device according to any one of claims 1 to 3 ;
A classifier that obtains probability of belonging to a known class and an unknown class for the data based on the parameters and the feature amount obtained by the learning device according to any one of claims 1 to 3 ;
A prediction device comprising: a feature extraction step of outputting feature quantities of input data using a feature extractor;
an identification step of obtaining probability of belonging to a known class and an unknown class for the data based on the feature amount using a plurality of classifiers;
an unknown class identification step of determining whether the data is an unknown class based on the acquired belonging probability;
an identification inconsistency evaluation step of outputting an identification inconsistency degree value indicating a difference in the respective attribution probabilities obtained by the plurality of classifiers for the data;
Using data that is not the unknown class and to which no teacher label has been assigned, the value of the identification discrepancy is reduced for the feature extractor, and the identification discrepancy for the plurality of classifiers is a learning step of iteratively learning the parameters of the feature extractor and the plurality of classifiers so as to increase the value;
has
further comprising a discriminative loss evaluation step for outputting a discriminative loss function value that is smaller as the degree of similarity between the attribution probability and a given teacher attribution probability of the data is higher for the data;
In the learning step, the feature extractor and the plurality of classifiers perform the identification using data to which a teacher label has been assigned and data that is an unknown class and to which no teacher label has been assigned. A learning method further comprising iteratively learning the parameters so as to reduce the value of the loss function . a feature extraction step of outputting feature quantities of input data using a feature extractor;
an identification step of obtaining probability of belonging to a known class and an unknown class for the data based on the feature amount using a plurality of classifiers;
an unknown class identification step of determining whether the data is an unknown class based on the acquired belonging probability;
an identification inconsistency evaluation step of outputting an identification inconsistency degree value indicating a difference in the respective attribution probabilities obtained by the plurality of classifiers for the data;
Using data that is not the unknown class and to which no teacher label has been assigned, the value of the identification discrepancy is reduced for the feature extractor, and the identification discrepancy for the plurality of classifiers is a learning step of iteratively learning the parameters of the feature extractor and the plurality of classifiers so as to increase the value;
has
In the unknown class identification step, if the number of times of iterative learning in the learning step is greater than a predetermined number of times, the learning method comprises making a determination based on past determination results . A program for operating a computer as the learning device according to claim 1 .

Images