Model learning device, model learning method, and program

Abstract

Provided is a model training technique for multi-label classification using data to which one or more complementary labels are assigned. This model training device comprises: a classification risk calculation unit that calculates the classification risk R(f) of a model f using a batch that is a subset of a set of data to which one or more complementary labels are assigned; and a model update unit that updates the model f using the classification risk R(f). When x is input data to the model f, -Y is a set of complementary labels assigned to the input data x, and f(x)k is the value of a model f(x) for a label k, the value of the classification risk R(f) decreases as the value f(x)k of the model f(x) for a label k that is not an element of the set -Y increases.

Description

[Technical field]

[0001] The present invention relates to machine learning techniques. [Background technology]

[0002] In recent years, machine learning has been used in the fields of acoustic event detection, image segmentation, image recognition, etc. The general steps of machine learning are as follows. (1) For each piece of data, one or more labels (called correct labels) that indicate the class to which the data belongs are assigned, and data with the correct labels is generated. By performing this process on a large amount of data, a set of data with the correct labels is generated. (2) Train the model using a set of data with correct labels. (3) Using the trained model, the class to which the input data belongs (called the correct class) is estimated.

[0003] Here, class estimation when there is one correct class for one piece of data is called multi-class classification. Class estimation when there are multiple possible correct classes for one piece of data is called multi-label classification. In multi-class classification, one correct label is assigned to one piece of data, while in multi-label classification, one or more correct labels are assigned to one piece of data.

[0004] In order to improve the accuracy of estimation of correct classes using a trained model generated according to the above procedure, data with more correct labels is required. However, as the number of classes to be estimated increases, similar classes appear, making it difficult to generate a large amount of data with correct labels accurately.

[0005] Therefore, Non-Patent Document 1 proposes the following procedure as machine learning for multi-label classification. (1) One or more labels (called complementary labels) are assigned to each piece of data to indicate classes to which the data does not belong, and data with complementary labels is generated. By performing this process on a large amount of data, a set of data with complementary labels is generated. As can be seen from the definition, complementary labels are labels that indicate that the data is not in the correct class. (2) Train the model using a set of complementary labeled data. (3) Using the trained model, the correct class is estimated for the input data.

[0006] Since it is generally easier to assign complementary labels to data than to assign correct labels to data, data to which more labels have been assigned can be easily generated. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] Keigo Wakayama, Shoichiro Saito, "Audio event detection based on complementary label learning," IEICE Technical Report EA2021-17, pp.77-82, 2021. Summary of the Invention [Problem to be solved by the invention]

[0008] In the technology of Non-Patent Document 1, data to which two or more supplementary labels are assigned in the above step (2) is converted into data to which one supplementary label is assigned, and then used for learning. Therefore, there is a problem in that the advantage that two or more supplementary labels are assigned to one piece of data during learning is not fully utilized.

[0009] Therefore, an object of the present invention is to provide a model learning technique for multi-label classification using data to which one or more complementary labels are assigned. [Means for solving the problem]

[0010] One aspect of the present invention is a model learning device including a classification risk calculation unit that calculates a classification risk R(f) of a model f using a batch that is a subset of a set of data to which one or more complementary labels have been assigned, and a model update unit that updates the model f using the classification risk R(f), - Let Y be the set of labels given to the input data x, and f(x) k Let be the value of the model f(x) for label k, and the classification risk R(f) is - The value of the model f(x) for label k that is not an element of Y k The larger it is, the smaller the value becomes.

[0011] One aspect of the present invention is a model learning device including a candidate labeled data set generation unit that generates a set of data to which one or more candidate labels have been assigned (hereinafter referred to as a candidate labeled data set) from a set of data to which one or more complementary labels have been assigned, a classification risk calculation unit that calculates a classification risk R(f) of a model f using a batch that is a subset of the candidate labeled data set, and a model update unit that updates the model f using the classification risk R(f), where x is input data to the model f, Y is a set of candidate labels assigned to the input data x, and f(x) k Let f(x) be the value of the model f(x) for label k, and the classification risk R(f) is the value f(x) of the model f(x) for label k, which is an element of the set Y. k The larger it is, the smaller the value becomes. Effect of the Invention

[0012] The present invention enables model training for multi-label classification using data that is annotated with one or more complementary labels. [Brief description of the drawings]

[0013] [Figure 1] FIG. 1 is a block diagram showing a configuration of a model learning device 100. [Diagram 2] 4 is a flowchart showing the operation of the model learning device 100. [Diagram 3]FIG. 2 is a block diagram showing a configuration of a model learning device 200. [Figure 4] 4 is a flowchart showing the operation of the model learning device 200. [Diagram 5] FIG. 2 is a diagram illustrating an example of a functional configuration of a computer that realizes each device according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] Hereinafter, an embodiment of the present invention will be described in detail. Note that components having the same functions are given the same reference numbers and duplicated explanations will be omitted.

[0015] Before describing each embodiment, the notation used in this specification will be described.

[0016] ^ (caret) represents a superscript. For example, x y^z Yes z is a superscript to x, and x y^z Yes z is a subscript to x. Also, _ (underscore) represents a subscript. For example, x y_z Yes z is a superscript to x, and x y_z Yes z is a subscript on x.

[0017] The superscripts "^" and "~" such as ^x and ~x for a certain letter x should actually be written directly above the "x", but due to the constraints of the description in the specification, they are written as ^x and ~x. <Technical background> In an embodiment of the present invention, a model for a multi-label classification problem is trained using data that is annotated with one or more complementary labels.

[0018] Let K be the number of classes into which data is classified, and let [K] = {1, ..., K} denote the set of labels. Correct labels are labels indicating that data belongs to class 1, ..., labels indicating that data belongs to class K. A set S is said to be a set of correct labels when set S is a subset of set [K], and the labels that are elements of set S are correct labels. Complementary labels are labels indicating that data does not belong to class 1, ..., labels indicating that data does not belong to class K. A set S is said to be a set of correct labels when set S is a subset of set [K], and the labels that are elements of set S are correct labels. A complementary label is a label indicating that data does not belong to class 1, ..., labels indicating that data does not belong to class K. - T is a set of complementary labels if the set - T is a subset of the set [K], - The labels that are elements of T are complementary labels. Here, a label that indicates that it does not belong to class k can be said to be a label that indicates that it belongs to a class other than k. Furthermore, the set of complementary labels - The complement of T, T=[K]- - T is called the set of candidate labels. Being an element of set T means that it is a correct candidate label. For example, when K=6, the set S of correct labels for data x is S={1, 3}, which means that data x belongs to class 1 and class 3. Similarly, the set of complementary labels for data x is - T - T={2, 4, 6} means that data x belongs to a class other than class 2, class 4, and class 6, and the set of candidate labels T for data x is T=[K]- - T={1, 3, 5} means that the data x may belong to class 1, class 3, or class 5.

[0019] The model f to be trained has the probability p k This is a function that outputs a K-dimensional vector with (k=1, …, K) as the kth element. Hereinafter, probability p k f(x) k Let us express it as follows.

[0020] Below, we will explain complementary label learning and candidate label learning. Here, complementary label learning refers to learning a model f using a set of data to which one or more complementary labels are assigned, and candidate label learning refers to learning a model f using a set of data to which one or more candidate labels are assigned. [Complementary label learning] In supplementary label learning, N is the batch size, x n (n=1, …, N) is the nth data in the batch. - Y n (n=1, …, N) is the data x n The classification risk R(f) of model f is calculated as the set of complementary labels assigned to

[0021]

number

[0022]

number

[0023] Instead of equation (1), the following equation may be used.

[0024]

number

[0025] (where ε is a given constant) In supplementary label learning, the model f is updated so that the classification risk R(f) becomes smaller. Note that when the classification risk R(f) in formula (1') is used, learning becomes more stable than when the classification risk R(f) in formula (1) is used. [Candidate label learning] In candidate label learning, N is the batch size, x n (n=1, …, N) is the nth data in the batch, Y n (n=1, …, N) is the data x n A set of candidate labels assigned to Z n(n=1, …, N) is the data x n As a set of correct labels, the classification risk R(f) of model f is calculated using the following formula:

[0026]

number

[0027] Instead of equation (3), the following equation may be used.

[0028]

number

[0029] (where ε is a given constant) In addition, set Z n Since it is impossible to know exactly what n is, we estimate it using a specific method and define it as the set Z n Used as.

[0030] In the candidate label learning, the model f is updated so that the classification risk R(f) becomes smaller. Note that when the classification risk R(f) in formula (3') is used, learning becomes more stable than when the classification risk R(f) in formula (3) is used. First Embodiment The model learning device 100 learns a model f using a set of data to which one or more supplementary labels are assigned (hereinafter, referred to as a supplementary-labeled data set). When learning a model used for acoustic event detection, the model f can be configured as a deep neural network (DNN) equipped with a self-attention mechanism described in Reference Non-Patent Document 1. The DNN can be, for example, a neural network including a convolutional neural network (CNN) that receives a log-mel spectrogram, which is a time-frequency representation of an acoustic signal, as input, and a transformer encoder. When learning a model used for image segmentation, the model f can be configured as a DNN equipped with a class activation map described in Reference Non-Patent Document 2. (Reference Non-Patent Document 1: Q. Kong et al., “Sound Event Detection of Weakly Labelled Data with CNN-Transformer and Automatic Threshold Optimization,” IEEE / ACM Transactions on Audio, Speech, and Language Processing, vol.28, pp.2450-2460, 2020.) (Reference Non-Patent Document 2: Y. Wang et al., “Self-supervised Equivariant Attention Mechanism for Weakly Supervised Semantic Segmentation,” CVPR 2020, pp.12275-12284, 2020.) Hereinafter, the model learning device 100 will be described with reference to Figs. 1 and 2. Fig. 1 is a block diagram showing the configuration of the model learning device 100. Fig. 2 is a flowchart showing the operation of the model learning device 100. As shown in Fig. 1, the model learning device 100 includes an initialization unit 110, a classification risk calculation unit 120, a model update unit 130, a termination condition determination unit 140, and a recording unit 190. The recording unit 190 is a component that appropriately records information necessary for the processing of the model learning device 100. The recording unit 190 records, for example, a complementary labeled data set in advance.

[0031] The operation of the model learning device 100 will be described with reference to FIG.

[0032] In S110, the initialization unit 110 sets initial values ​​of the model f. For example, the initialization unit 110 sets initial values ​​of parameters of the DNN that constitutes the model f.

[0033] In S120, the classification risk calculation unit 120 calculates the classification risk R(f) of the model f using a batch that is a subset of the complementary labeled data set recorded in the recording unit 190. The classification risk calculation unit 120 calculates the classification risk R(f) by, for example, equations (1) and (2), or equations (1') and (2). In other words, the classification risk R(f) is calculated by: - Let Y be the set of labels given to the input data x, and f(x) k Let be the value of the model f(x) for label k, and let - The value of the model f(x) for label k that is not an element of Y k The calculation is performed using a function whose value becomes smaller as

[0034] In S130, the model update unit 130 updates the model f by using the classification risk R(f) calculated in S120. The model update unit 130 updates the parameters of the DNN constituting the model f, for example, so that the classification risk R(f) becomes smaller.

[0035] In S140, if a predetermined termination condition is satisfied, the termination condition determination unit 140 terminates the process by treating the model f obtained in the process of S130 as a trained model, and otherwise returns to the process of S120. For example, the termination condition may be a condition as to whether or not the upper limit of the number of model updates has been reached.

[0036] According to an embodiment of the present invention, model learning for multi-label classification using data to which one or more supplementary labels are assigned is possible. By learning using data to which one or more supplementary labels are assigned as is without converting it to data to which one supplementary label is assigned, more accurate multi-label classification is possible. <Second embodiment> The model learning device 200 learns a model f using a set of data to which one or more complementary labels have been added (hereinafter, referred to as a complementary labeled data set).

[0037] Hereinafter, the model learning device 200 will be described with reference to Figs. 3 and 4. Fig. 3 is a block diagram showing the configuration of the model learning device 200. Fig. 4 is a flowchart showing the operation of the model learning device 200. As shown in Fig. 3, the model learning device 200 includes a candidate labeled data set generating unit 205, an initialization unit 110, a classification risk calculation unit 220, a model updating unit 130, a termination condition determination unit 140, and a recording unit 190. The recording unit 190 is a component that appropriately records information necessary for the processing of the model learning device 200.

[0038] The operation of model learning device 200 will be described with reference to FIG.

[0039] In S205, the candidate labeled data set generating unit 205 generates, from the complementary labeled data set recorded in the recording unit 190, a set of data to which one or more candidate labels have been assigned (hereinafter, referred to as a candidate labeled data set).

[0040] In S110, the initialization unit 110 sets an initial value of the model f.

[0041] In S220, the classification risk calculation unit 220 calculates the classification risk R(f) of the model f using a batch that is a subset of the candidate labeled data set generated in S205. The classification risk calculation unit 220 calculates the classification risk R(f) by, for example, formula (3) or formula (3'). That is, the classification risk R(f) is expressed as follows: x is the input data to the model f, Y is the set of candidate labels assigned to the input data x, f(x) k Let f(x) be the value of the model f(x) for label k, and let f(x) be the value of the model f(x) for label k, which is an element of the set Y. k The calculation is performed using a function whose value becomes smaller as

[0042] In S130, the model update unit 130 updates the model f by using the classification risk R(f) calculated in S220.

[0043] In S140, if a predetermined termination condition is satisfied, the termination condition determination unit 140 terminates the process by treating the model f obtained in the process of S130 as a trained model, and otherwise returns to the process of S220.

[0044] According to an embodiment of the present invention, model learning for multi-label classification using data to which one or more supplementary labels are assigned is possible. By converting data to which one or more supplementary labels are assigned into data to which one or more candidate labels are assigned and then using the data as is for learning, more accurate multi-label classification is possible. <Additional Notes> 5 is a diagram showing an example of the functional configuration of a computer that realizes each of the above-mentioned devices (i.e., each node). The processing in each of the above-mentioned devices can be implemented by having a recording unit 2020 load a program for causing a computer to function as each of the above-mentioned devices, and having the control unit 2010, input unit 2030, output unit 2040, etc. operate.

[0045] The device of the present invention has, as a single hardware entity, an input section to which a keyboard or the like can be connected, an output section to which a liquid crystal display or the like can be connected, a communication section to which a communication device (e.g., a communication cable) capable of communicating with the outside of the hardware entity can be connected, a CPU (which may also have a central processing unit, cache memory, registers, etc.), memories such as RAM and ROM, an external storage device such as a hard disk, and buses connecting the input section, output section, communication section, CPU, RAM, ROM, and external storage device so that data can be exchanged between them. If necessary, the hardware entity may also be provided with a device (drive) capable of reading and writing a recording medium such as a CD-ROM. A physical entity equipped with such hardware resources is, for example, a general-purpose computer.

[0046] The external storage device of the hardware entity stores the programs required to realize the above-mentioned functions and the data required in the processing of these programs (not limited to the external storage device, but for example the programs may be stored in a ROM, which is a read-only storage device). Data obtained by the processing of these programs is stored appropriately in the RAM, the external storage device, etc.

[0047] In the hardware entity, each program stored in an external storage device (or ROM, etc.) and the data required for processing each program are loaded into memory as necessary, and interpreted, executed, and processed by the CPU as appropriate, so that the CPU realizes a predetermined function (each component represented as the above, "unit," "means," etc.).

[0048] The present invention is not limited to the above-described embodiment, and can be modified as appropriate without departing from the spirit of the present invention. In addition, the processes described in the above-described embodiment are not limited to being executed in chronological order according to the order described, but may be executed in parallel or individually depending on the processing capacity of the device that executes the processes or as necessary.

[0049] As described above, when the processing functions of the hardware entities (the devices of the present invention) described in the above embodiments are realized by a computer, the processing contents of the functions that the hardware entities should have are described by a program. Then, by executing this program on a computer, the processing functions of the hardware entities are realized on the computer.

[0050] The program describing the processing contents can be recorded on a computer-readable recording medium. The computer-readable recording medium may be, for example, a magnetic recording device, an optical disk, a magneto-optical recording medium, a semiconductor memory, or any other suitable recording medium. Specifically, for example, a hard disk drive, a flexible disk, a magnetic tape, or the like can be used as the magnetic recording device, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only Memory), a CD-R (Recordable) / RW (ReWritable), or the like can be used as the optical disk, an MO (Magneto-Optical disc), or the like can be used as the magneto-optical recording medium, and an EEP-ROM (Electronically Erasable and Programmable-Read Only Memory) or the like can be used as the semiconductor memory.

[0051] The program may be distributed, for example, by selling, transferring, lending, etc. portable recording media such as DVDs and CD-ROMs on which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of a server computer and transferring the program from the server computer to other computers via a network.

[0052] A computer that executes such a program, for example, first stores the program recorded on a portable recording medium or the program transferred from a server computer in its own storage device. Then, when executing a process, the computer reads the program stored in its own storage device and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from the portable recording medium and execute a process according to the program, or may execute a process according to the received program each time a program is transferred from the server computer to the computer. The server computer may not transfer the program to the computer, but may execute the above-mentioned process by a so-called ASP (Application Service Provider) type service that realizes the processing function only by issuing an execution instruction and obtaining the result. Note that the program in this embodiment includes information used for processing by an electronic computer that is equivalent to a program (data that is not a direct instruction to the computer but has a nature that specifies the processing of the computer, etc.).

[0053] In addition, in this embodiment, a hardware entity is configured by executing a specific program on a computer, but at least a part of the processing contents may be realized by hardware.

[0054] The foregoing description of the embodiments of the invention has been presented for purposes of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise form disclosed. Modifications and variations are possible in light of the above teachings. The embodiments have been chosen and depicted to provide a best illustration of the principles of the invention and to enable those skilled in the art to utilize the invention in various embodiments and with various modifications as may be suitable for the practical use contemplated. All such modifications and variations are within the scope of the invention as defined by the appended claims interpreted in accordance with the breadth to which they are fairly, legally, and equitably entitled.

Claims

1. A classification risk calculation unit that calculates a classification risk R(f) of a model f using a batch that is a subset of a set of data to which one or more supplementary labels have been assigned; a model updating unit that updates the model f using the classification risk R(f); A model learning device comprising: Let x be the input data to model f. - Y is the set of labels given to the input data x, and f(x) is k Let be the value of the model f(x) for label k, Let N be the batch size, xn (n=1, …, N) be the nth data in the batch, - Y n (n=1, …, N) be the set of complementary labels assigned to data xn, and K be the number of labels. The classification risk R(f) is calculated by the following formula: [006] [0070] Model learning device.

2. a candidate labeled data set generation unit that generates a set of data to which one or more candidate labels have been assigned (hereinafter referred to as a candidate labeled data set) from a set of data to which one or more complementary labels have been assigned; A classification risk calculation unit that calculates a classification risk R(f) of a model f using a batch that is a subset of the candidate labeled data set; a model updating unit that updates the model f using the classification risk R(f); A model learning device comprising: Let x be the input data to the model f, Y be the set of candidate labels given to the input data x, and f(x). k Let be the value of the model f(x) for label k, Let N be the batch size, xn (n=1, …, N) be the nth data in the batch, Y n (n=1, …, N) be the set of candidate labels assigned to data xn, and Z n (n=1, …, N) be the set of correct labels for data xn. The classification risk R(f) is calculated by the following formula: [0080] Model learning device.

3. A classification risk calculation step in which the model learning device calculates a classification risk R(f) of a model f using a batch that is a subset of a set of data to which one or more supplementary labels have been assigned; a model updating step in which the model learning device updates the model f using the classification risk R(f); A model learning method comprising: Let x be the input data to model f. - Y is the set of labels given to the input data x, and f(x) is k Let be the value of the model f(x) for label k, Let N be the batch size, xn (n=1, …, N) be the nth data in the batch, - Y n (n=1, …, N) be the set of complementary labels assigned to data xn, and K be the number of labels. The classification risk R(f) is calculated by the following formula: [0090] [0010] Model training methods.

4. a candidate labeled data set generating step in which the model learning device generates a set of data to which one or more candidate labels have been assigned (hereinafter referred to as a candidate labeled data set) from a set of data to which one or more complementary labels have been assigned; A classification risk calculation step in which the model learning device calculates a classification risk R(f) of a model f using a batch that is a subset of the candidate labeled data set; a model updating step in which the model learning device updates the model f using the classification risk R(f); A model learning method comprising: Let x be the input data to the model f, Y be the set of candidate labels given to the input data x, and f(x). k Let be the value of the model f(x) for label k, Let N be the batch size, xn (n=1, …, N) be the nth data in the batch, Y n (n=1, …, N) be the set of candidate labels assigned to data xn, and Z n (n=1, …, N) be the set of correct labels for data xn. The classification risk R(f) is calculated by the following formula: ##EQU00011## Model training methods.

5. A program for causing a computer to function as the model learning device according to claim 1 or 2.

Images