JP7540595B2 - Model learning device, model learning method, and program - Google Patents

Description

The present invention relates to machine learning technology.

In recent years, machine learning has been used in fields such as acoustic event detection, image segmentation, and image recognition. The general steps of machine learning are described below.

(1) A label (called a correct label) is assigned to each piece of data, indicating the class to which the data belongs, and data with the correct label is generated. By performing this process on many pieces of data, a set of data with 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.

When trying to improve the accuracy of correct class estimation using a trained model generated according to the above procedure, data with more correct labels is required. Also, when trying to increase the number of classes to which the data to be estimated belongs, data with more correct labels is required. However, the task of assigning correct labels is very time-consuming, and it is difficult to generate large amounts of data with correct labels.

Therefore, non-patent document 1 proposes the following machine learning procedure.

(1) A label (called a complementary label) is assigned to each piece of data, indicating a class to which the data does not belong, and data is generated that has been assigned the incorrect complementary label. By performing this process on many pieces of data, a set of data that has been assigned complementary labels is generated.

(2) Train the model using a set of data with complementary labels.

(3) Using the trained model, the class to which the input data belongs (called the correct class) is estimated.

Since it is less time-consuming to assign complementary labels to data than to assign correct labels to data, it is possible to generate data with more labels. In addition, the estimation accuracy of the correct class using a trained model generated according to the procedure in Non-Patent Document 1 is comparable to the estimation accuracy of the correct class using a trained model generated according to a general procedure.

T. Ishida et al., “Complementary-Label Learning for Arbitrary Losses and Models,” ICML 2019, pp.2971-2980, 2019.

The technology in Non-Patent Document 1 targets problems where a single piece of data belongs to only one class (hereafter referred to as a multi-class classification problem), and cannot handle problems where a single piece of data belongs to more than one class (i.e., there may be two or more classes) (hereafter referred to as a multi-label classification problem).

Therefore, the present invention aims to provide a model learning technology using complementary labels for multi-label classification problems.

（ただし、Kはデータを分類するクラスの数、lossは１個以上の正解ラベルが付与されたデータの集合を用いてモデルを学習する場合に用いる損失関数）、リスク^-R(g:^-loss)を用いて、モデルを更新するモデル更新部と、を含む。 One aspect of the present invention includes a training data generation unit that generates a set of data to which one supplementary label is assigned (hereinafter referred to as a training data set) from a set of data to which one or more supplementary labels are assigned; and a first risk calculation unit that calculates a risk ^- R(g: ^-loss ) of a decision function g related to a loss function ^- loss, which is calculated by the following formula, using a batch that is a subset of the training data set;

(where K is the number of classes into which data is classified, and loss is a loss function used when training a model using a set of data to which one or more correct labels are assigned), and a model update unit that updates the model using risk ^- R(g: ^- loss).

（ただし、Kはデータを分類するクラスの数、lossは１個以上の正解ラベルが付与されたデータの集合を用いてモデルを学習する場合に用いる損失関数）、１個以上の正解ラベルが付与されたデータの集合の部分集合であるバッチを用いて、損失関数lossに関する決定関数gのリスクR(g:loss)を計算する第２リスク計算部と、リスク^-R(g:^-loss)とリスクR(g:loss)から、R(g)=α^-R(g:^-loss)+(1-α)(g:loss)（ただし、αは0<α<1を満たす定数）により、リスクR(g)を計算する第３リスク計算部と、リスクR(g)を用いて、モデルを更新するモデル更新部と、を含む。 One aspect of the present invention includes a training data generation unit that generates a set of data to which one supplementary label is assigned (hereinafter referred to as a training data set) from a set of data to which one or more supplementary labels are assigned; and a first risk calculation unit that calculates a risk ^- R(g: ^-loss ) of a decision function g related to a loss function ^- loss, which is calculated by the following formula, using a batch that is a subset of the training data set;

(where K is the number of classes for classifying data, and loss is a loss function used when training a model using a set of data to which one or more correct labels have been assigned), a second risk calculation unit that calculates a risk R(g:loss) of a decision function g related to the loss function loss using a batch that is a subset of the set of data to which one or more correct labels have been assigned; a third risk calculation unit that calculates a risk R(g) from the risk ^-R (g: ^-loss ) and the risk R(g:loss) by R(g)=α ^- R(g: ^-loss )+(1-α)(g:loss) (where α is a constant satisfying 0<α<1); and a model update unit that updates the model using the risk R(g).

According to the present invention, it becomes possible to learn a model using complementary labels for multi-label classification problems.

FIG. 1 is a block diagram showing a configuration of a model learning device 100. 4 is a flowchart showing the operation of the model learning device 100. FIG. 2 is a block diagram showing a configuration of a model learning device 200. 4 is a flowchart showing the operation of the model learning device 200. 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.

Hereinafter, an embodiment of the present invention will be described in detail. Components having the same functions are given the same numbers, and duplicate explanations will be omitted.

Before describing each embodiment, we will explain the notation used in this specification.

^ (caret) represents a superscript. For example, x ^{y^z} means that y ^z is a superscript to x, and x _y^z means that y ^z is a subscript to x. _ (underscore) represents a subscript. For example, x ^y_z means that y _z is a superscript to x, and x _{y_z} means that y _z is a subscript to x.

In addition, superscripts such as "^" and "~" for a certain letter x, such as ^x and ~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 with one or more supplementary labels. Note that the embodiment of the present invention can also be used to train a model for a multi-class classification problem using data with one or more supplementary labels.

In the following, the number of classes to which data may belong, i.e., the number of classes into which data is classified, is defined as K, and the set of correct labels is [K] = {1, ..., K}. Here, correct labels are labels indicating that data belongs to class 1, ..., and labels indicating that data belongs to class K, and are represented as 1, ..., K, respectively.

Consider K complementary labels. Here, the K complementary labels are labels indicating that the class does not belong to class 1, ..., and labels indicating that the class does not belong to class K, and are represented as ^-1 , ..., ^-K , respectively. The set of complementary labels is represented as [ ^-K ]={ ^- 1, ..., ^-K }.

In the embodiment of the present invention, data to which one or more complementary labels are assigned is handled as follows. For data to which M complementary labels are assigned, where M is 2 or more, M pieces of data to which one complementary label is assigned are generated from the data. This can be explained using the symbols described below as "data (x _i , ( ^-y ₁ , ..., ^-y _M )) to which M complementary labels are assigned (where x _i ∈χ, ^-y ₁ , ..., ^-y _M ∈[ ^-K ]) are generated as data (x _i , ^-y ₁ ), ..., (x _i , ^-y _M ) to which one complementary label is assigned."

Below, we will explain correct label learning and supplementary label learning in detail. Correct label learning here means learning a model using a set of data to which one or more correct labels have been assigned, and supplementary label learning means learning a model using a set of data to which one or more supplementary labels have been assigned.

また、リスクR(g:loss)は、次式で表すこともできる。

マルチラベル分類問題に対するモデルを学習する場合、損失関数lossとして、以下のバイナリクロスエントロピーやマルチラベルソフトマージンを用いることができる。ここで、y_kはクラスkが存在する場合は1、それ以外の場合は0を表すものとする。 [Correct label learning]
Let χ be the set of data, g:χ→R ^K be the decision function, and g _k be the k-th element of the decision function g. Let D be the distribution on χ×[K] (where the random variables of distribution D are represented as (X, Y) to D), {P _k } _k=1 ^K (where P _k =P(X|Y=k)), {π _k } _k=1 ^K (where π _k =P(Y=k)), and loss:[K]×R ^K →R ₊ be the loss function for learning correct labels. The loss function loss, the risk R(g:loss) of the decision function g on distribution D, is expressed by the following equation.

In addition, the risk R(g:loss) can also be expressed by the following formula:

When training a model for a multi-label classification problem, the following binary cross entropy or multi-label soft margin can be used as the loss function, where y _k is 1 if class k exists and 0 otherwise.

（マルチラベルソフトマージン）

なお、多クラス分類問題に対するモデルを学習する場合、損失関数lossとして、ソフトマックスクロスエントロピーを用いることができる。 (Binary Cross Entropy)

(Multi-label soft margin)

When training a model for a multi-class classification problem, softmax cross entropy can be used as the loss function.

また、リスク^-R(g:^-loss)は、次式で表すこともできる。

損失関数^-lossは、損失関数lossを用いた次式で計算される。

なお、１個以上の正解ラベルが付与されたデータの集合と１個以上の補ラベルが付与されたデータの集合とを用いてモデルを学習することもできる。この場合、決定関数gのリスクR(g)は、次式を用いて計算するとよい。

ただし、αは0<α<1を満たす定数である。 [Complementary label learning]
^{Let -D} be a distribution on χ×[ ^-K ] (where the random variable in distribution ^-D is represented as (X, ^-Y )～ ^-D ), { ^-Pk } _k=1K ^{( where -Pk} ₌ P(X| ^-Y = _k )), { ^-πk } _{k =1K} ⁽ where ^-πk =P( _-Y =k)), ^and -loss:[ ^- K]× ^RK →R ₊ be the loss function for complementary label learning. Then, the loss function ^-loss , ^the risk ^-R (g: ^-loss ) of the decision function g on distribution ^-D is expressed by the following equation.

The risk ^−R (g: ^−loss ) can also be expressed as follows:

The loss function ^- loss is calculated using the loss function loss as follows:

It is also possible to train a model using a set of data to which one or more correct labels have been assigned and a set of data to which one or more complementary labels have been assigned. In this case, the risk R(g) of the decision function g may be calculated using the following formula:

Here, α is a constant that satisfies 0<α<1.

First Embodiment
Below, the model learning device 100 will be described with reference to Figures 1 and 2. Figure 1 is a block diagram showing the configuration of the model learning device 100. Figure 2 is a flowchart showing the operation of the model learning device 100. As shown in Figure 1, the model learning device 100 includes a learning data generation unit 110, a first 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 operation of the model learning device 100 will be explained with reference to Figure 2.

At S110, the training data generation unit 110 generates a set of data to which one complementary label has been assigned (hereinafter referred to as a training data set) from a set of data to which one or more complementary labels have been assigned (hereinafter referred to as an input complementary labeled data set).

（ただし、Kはデータを分類するクラスの数、lossは１個以上の正解ラベルが付与されたデータの集合を用いてモデルを学習する場合に用いる損失関数）
Ｓ１３０において、モデル更新部１３０は、Ｓ１２０で計算したリスク^-R(g:^-loss)を用いて、モデルを更新する。具体的には、モデル更新部１３０は、リスク^-R(g:^-loss)を最小化するように、モデルを更新する。音響イベント検知に用いるモデルを学習する場合、モデルは、参考非特許文献１に記載の自己注意機構を備えたDNNモデルとすることができる。また、画像セグメンテーションに用いるモデルを学習する場合、モデルは、参考非特許文献２に記載のクラスアクティベーションマップを備えたDNNモデルとすることができる。 In S120, the first risk calculation unit 120 calculates a risk ^−R (g:-loss) of a decision function g related to a loss function ^−loss calculated by the following formula, using a batch which is a subset of the ^training data set generated in S110.

(where K is the number of classes into which data is classified, and loss is the loss function used when training a model using a set of data with one or more correct labels.)
In S130, the model update unit 130 updates the model using the risk ^- R(g: ^-loss ) calculated in S120. Specifically, the model update unit 130 updates the model so as to minimize the risk ^- R(g: ^-loss ). When training a model for use in acoustic event detection, the model may be a DNN model equipped with a self-attention mechanism as described in Reference Non-Patent Document 1. When training a model for use in image segmentation, the model may be a DNN model equipped with a class activation map as described in Reference Non-Patent Document 2.

(Reference Non-Patent Document 1: Q. Kong et al., “Sound Event Detection of Weakly Labeled 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.)
In S140, if a predetermined termination condition is satisfied, the termination condition determination unit 140 terminates the process by treating the model obtained in the process of S130 as a trained model, and otherwise returns to the process of S120. The termination condition may be, for example, whether or not the upper limit of the number of model updates has been reached.

(Modification)
The above-mentioned processes in steps S110 to S140 may be carried out as follows.

In S110, the training data generation unit 110 generates a batch that is a subset of the input labeled data set from a set of data to which one or more complementary labels have been assigned (hereinafter referred to as the input labeled data set), and generates a set of data to which one complementary label has been assigned from the batch (hereinafter referred to as the training data set).

In S120, the first risk calculation unit 120 calculates the risk ^−R (g: ^-loss ) of the decision function g related to the loss function ^−loss calculated by the formula (1) using the training data set generated in S110.

In S130, the model update unit 130 updates the model using the risk ^−R (g: ^−loss ) calculated in S120.

In S140, if a specified termination condition is satisfied, the termination condition determination unit 140 terminates the processing by treating the model obtained in the processing of S130 as the trained model, and otherwise returns to the processing of S110.

According to an embodiment of the present invention, it is possible to learn a model using a complementary label for a multi-label classification problem. By using data with a complementary label, it is possible to learn using a larger amount of data, thereby improving the accuracy of estimation using a trained model and enabling estimation for a larger number of classes.

Second Embodiment
Hereinafter, model learning device 200 will be described with reference to Figs. 3 and 4. Fig. 3 is a block diagram showing the configuration of model learning device 200. Fig. 4 is a flowchart showing the operation of model learning device 200. As shown in Fig. 3, model learning device 200 includes a learning data generation unit 110, a first risk calculation unit 120, a second risk calculation unit 220, a third risk calculation unit 230, a model update unit 240, a termination condition determination unit 140, and a recording unit 190. Recording unit 190 is a component that appropriately records information necessary for the processing of model learning device 200.

The operation of the model learning device 200 will be explained with reference to Figure 4.

At S110, the training data generation unit 110 generates a set of data to which one complementary label has been assigned (hereinafter referred to as a training data set) from a set of data to which one or more complementary labels have been assigned (hereinafter referred to as an input complementary labeled data set).

（ただし、Kはデータを分類するクラスの数、lossは１個以上の正解ラベルが付与されたデータの集合を用いてモデルを学習する場合に用いる損失関数）
Ｓ２２０において、第２リスク計算部２２０は、１個以上の正解ラベルが付与されたデータの集合（以下、入力正解ラベル付きデータ集合という）の部分集合であるバッチを用いて、損失関数lossに関する決定関数gのリスクR(g:loss)を計算する。 In S120, the first risk calculation unit 120 calculates a risk ^−R (g:-loss) of a decision function g related to a loss function ^−loss calculated by the following formula, using a batch which is a subset of the ^training data set generated in S110.

(where K is the number of classes into which data is classified, and loss is the loss function used when training a model using a set of data with one or more correct labels.)
In S220, the second risk calculation unit 220 calculates the risk R(g:loss) of the decision function g related to the loss function loss using a batch that is a subset of a set of data to which one or more correct labels have been assigned (hereinafter referred to as the input correct-labeled data set).

In S230, the third risk calculation unit 230 calculates a risk R(g) from the risk ^-R (g: ^-loss ) calculated in S120 and the risk R(g:loss) calculated in S130 by R(g)=α ^- R(g: ^-loss )+(1-α)(g:loss) (where α is a constant satisfying 0<α<1).

In S240, the model update unit 240 updates the model using the risk R(g) calculated in S230. Specifically, the model update unit 240 updates the model so as to minimize the risk R(g). Also, as in the first embodiment, when training a model to be used for acoustic event detection, the model described in Reference Non-Patent Document 1 can be used, and when training a model to be used for image segmentation, the model described in Reference Non-Patent Document 2 can be used.

In S140, if a specified termination condition is satisfied, the termination condition determination unit 140 terminates the processing by treating the model obtained in the processing of S240 as the trained model, and otherwise returns to the processing of S120 and S220.

(Modification)
The above-mentioned processes in steps S110 to S140 may be carried out as follows.

In S110, the training data generation unit 110 generates a batch that is a subset of the input labeled data set from a set of data to which one or more complementary labels have been assigned (hereinafter referred to as the input labeled data set), and generates a set of data to which one complementary label has been assigned (hereinafter referred to as the training data set) from the batch.

In S120, the first risk calculation unit 120 calculates the risk ^−R (g: ^-loss ) of the decision function g related to the loss function ^−loss calculated by the formula (2), using the training data set generated in S110.

In S220, the second risk calculation unit 220 calculates the risk R(g:loss) of the decision function g related to the loss function loss using a batch, which is a subset of a set of data to which one or more correct answer labels have been assigned (hereinafter referred to as the input labeled data set).

In S230, the third risk calculation unit 230 calculates a risk R(g) from the risk ^-R (g: ^-loss ) calculated in S120 and the risk R(g:loss) calculated in S130 by R(g)=α ^- R(g: ^-loss )+(1-α)(g:loss) (where α is a constant satisfying 0<α<1).

In S240, the model update unit 240 updates the model using the risk R(g) calculated in S230.

In S140, if a specified termination condition is satisfied, the termination condition determination unit 140 terminates the processing by treating the model obtained in the processing of S240 as the trained model, and otherwise returns to the processing of S110 and S220.

According to an embodiment of the present invention, it is possible to learn a model using a complementary label for a multi-label classification problem. By using data with a complementary label, it is possible to learn using a larger amount of data, thereby improving the accuracy of estimation using a trained model and enabling estimation for a larger number of classes.

<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 the recording unit 2020 load a program for causing the computer to function as each of the above-mentioned devices, and having the control unit 2010, input unit 2030, output unit 2040, etc. operate.

The device of the present invention, for example as a single hardware entity, has an input section to which a keyboard or the like can be connected, an output section to which an LCD display or the like can be connected, a communication section to which a communication device (for example 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 these 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 recording media such as a CD-ROM. An example of a physical entity equipped with such hardware resources is a general-purpose computer.

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

In a hardware entity, each program stored in an external storage device (or ROM, etc.) and the data required to process each program are loaded into memory as needed, and interpreted, executed, and processed by the CPU as appropriate. As a result, the CPU realizes a specified function (each component represented as the above, ... unit, ... means, etc.).

The present invention is not limited to the above-described embodiments, and appropriate modifications can be made without departing from the spirit of the present invention. Furthermore, the processes described in the above embodiments are not limited to being executed chronologically in the order described, but may be executed in parallel or individually depending on the processing capacity of the device executing the processes or as necessary.

As mentioned 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.

The program describing the processing contents can be recorded on a computer-readable recording medium. Examples of computer-readable recording media include magnetic recording devices, optical disks, magneto-optical recording media, and semiconductor memories. Specifically, for example, a hard disk drive, a flexible disk, a magnetic tape, etc. can be used as a 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), etc. can be used as an optical disk; an MO (Magneto-Optical disc) can be used as a magneto-optical recording medium; and an EEP-ROM (Electronically Erasable and Programmable-Read Only Memory) can be used as a semiconductor memory.

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.

A computer that executes such a program, for example, first stores in its own storage device the program recorded on a portable recording medium or the program transferred from a server computer. Then, when executing a process, the computer reads the program stored in its own storage device and executes the process according to the read program. As another execution form of this program, the computer may read the program directly from the portable recording medium and execute the process according to the program, or may execute the process according to the received program each time a program is transferred from the server computer to this computer. In addition, the server computer may not transfer the program to this 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 results. Note that the program in this embodiment includes information used for processing by an electronic computer that is equivalent to a program (such as data that is not a direct command to the computer but has properties that specify the processing of the computer).

In addition, in this embodiment, a hardware entity is configured by executing a specific program on a computer, but at least a portion of these processing contents may also be realized by hardware.

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 (4)

a training data generation unit that generates a set of data to which one complementary label is assigned (hereinafter referred to as a training data set) from a set of data to which one or more complementary labels are assigned;
A first risk calculation unit that calculates a risk ^- R(g: ^-loss ) of a decision function g regarding a loss function ^- loss calculated by the following formula using a batch that is a subset of the learning data set;

(where K is the number of classes into which data is classified, and loss is the loss function used when training a model using a set of data with one or more correct labels.)
A second risk calculation unit that calculates a risk R(g:loss) of a decision function g related to a loss function loss using a batch that is a subset of a set of data to which one or more correct answer labels are assigned;
a third risk calculation unit that calculates risk R(g) from risk ^- R(g: ^- loss) and risk R(g:loss) by R(g)=α ^- R(g: ^- loss)+(1-α) R (g:loss) (where α is a constant satisfying 0<α<1);
a model updating unit that updates the model using the risk R(g);
A model learning device comprising: a learning data generating step in which the model learning device generates a set of data to which one supplementary label is assigned (hereinafter referred to as a learning data set) from a set of data to which one or more supplementary labels are assigned;
A first risk calculation step in which the model learning device calculates a risk ^- R(g: ^-loss ) of a decision function g regarding a loss function ^- loss calculated by the following formula, using a batch that is a subset of the learning data set;

(where K is the number of classes into which data is classified, and loss is the loss function used when training a model using a set of data with one or more correct labels.)
A second risk calculation step in which the model learning device calculates a risk R(g:loss) of a decision function g related to a loss function loss using a batch that is a subset of a set of data to which one or more correct answer labels are assigned;
a third risk calculation step in which the model learning device calculates a risk R(g) from risk ^-R (g: ^-loss ) and risk R(g:loss) by R(g)=α ^- R(g: ^-loss )+(1-α) R (g:loss) (where α is a constant satisfying 0<α<1);
a model updating step in which the model learning device updates the model using the risk R(g);
A model training method including: The model learning method according to claim 2 ,
A model training method characterized in that the loss function is binary cross entropy or multi-label soft margin. A program for causing a computer to execute the model learning method according to claim 2 or 3 .

Images