Information processing device, classification model generation method, teacher data generation method, learning program, and teacher data generation program

Abstract

To generate a classification model adapted to a prescribed domain without performing work for making correspondence with a correct label.SOLUTION: An information processing device (1) includes a classification model acquisition unit (102) that acquires a first classification model (111) used for classifying data having deviation in classification, and a first learning unit (103) that updates the first classification model (111) using target data including no correct label. The first learning unit (103) updates the first classification model (111) so as to, when a prescribed number of target data sets are classified into a plurality of classes by the first classification model (111), make the numbers of target data sets classified into the respective classes close to a prescribed ratio.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to an information processing apparatus that generates a classification model and the like.

Background Art

[0002] Domain adaptation techniques for adapting a classification model generated by machine learning using data in a certain domain to another domain have been conventionally known. For example, Patent Document 1 below describes generating a first model for identifying the types of waste from first image data obtained by imaging inside a garbage pit storing waste. And Patent Document 1 describes performing additional learning of the first model using second image data obtained by imaging inside a garbage pit of another facility to generate a second model adapted to the other facility.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In the technique of Patent Document 1, when generating the second model, it is necessary to visually identify the types of waste shown in the second image data, label them by type, and generate teacher data. Since such work requires a great deal of cost, a technique for automatically labeling is desired.

[0005] However, if the accuracy of the automatically assigned labels is low, the classification accuracy of the classification model in a predetermined domain cannot be improved. That is, it is not easy to generate a classification model adapted to a predetermined domain without performing the work of associating correct labels with the target data. This is a problem that occurs not only in classification using images but also when classifying arbitrary data by a machine-learned model.

[0006] One aspect of the present invention aims to realize an information processing apparatus or the like that enables generation of a classification model adapted to a predetermined domain without performing an operation of associating correct labels with target data.

Means for Solving the Problems

[0007] To solve the above problems, an information processing apparatus according to one aspect of the present invention includes a classification model acquisition unit that acquires a classification model that is a classification model to be adapted to a predetermined domain and is used for classifying data with a bias in classification, and a learning unit that updates the classification model using target data that belongs to the domain and does not include correct labels. The learning unit updates the classification model so that the number of the target data classified into each class when a predetermined number of the target data are classified into a plurality of classes by the classification model approaches a predetermined ratio obtained in advance according to the bias.

[0008] To solve the above problems, another information processing apparatus according to one aspect of the present invention includes a classification unit that classifies target data belonging to a predetermined domain, and a teacher data generation unit that generates teacher data for generating a classification model used for classifying data with a bias in classification by associating the result of the classification with the target data as a correct label. The classification unit classifies the target data using a classification model generated by machine learning so that the number of the target data classified into each class when a predetermined number of the target data are classified into a plurality of classes approaches a predetermined ratio obtained in advance according to the bias.

[0009] In order to solve the above problems, a method for generating a classification model according to an aspect of the present invention is a method for generating a classification model executed by at least one information processing device, including: an acquisition step of acquiring a classification model that is a classification model to be adapted to a predetermined domain and is used for classifying data with a bias in classification; and an update step of updating the classification model using target data that belongs to the domain and does not include correct labels. In the update step, the classification model is updated so that the number of pieces of target data classified into each class when a predetermined number of pieces of target data are classified into a plurality of classes by the classification model approaches a predetermined ratio obtained in advance according to the bias.

[0010] In order to solve the above problems, a method for generating teacher data according to an aspect of the present invention is a method for generating teacher data used for generating a classification model used for classifying data with a bias in classification, which is executed by at least one information processing device. The method includes: a classification step of classifying target data using a classification model generated by machine learning so that the number of pieces of target data classified into each class approaches a predetermined ratio obtained in advance according to the bias when a predetermined number of pieces of target data belonging to a predetermined domain are classified into a plurality of classes; and a teacher data generation step of generating teacher data by associating the result of the classification with the target data as correct labels.

Advantages of the Invention

[0011] According to an aspect of the present invention, it is possible to generate a classification model adapted to a predetermined domain without performing the operation of associating correct labels.

Brief Description of the Drawings

[0012]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Mode for Carrying Out the Invention

[0013] 〔Embodiment 1〕 (Configuration of Information Processing Apparatus 1) FIG. 1 is a block diagram showing an example of the main configuration of the information processing apparatus 1 according to the present embodiment. As shown in the figure, the information processing apparatus 1 includes a control unit 10 that comprehensively controls each part of the information processing apparatus 1, and a storage unit 11 that stores various data used by the information processing apparatus 1. Further, the information processing apparatus 1 includes a communication unit 12 for the information processing apparatus 1 to communicate with other devices, an input unit 13 for receiving inputs of various data to the information processing apparatus 1, and an output unit 14 for the information processing apparatus 1 to output various data.

[0014] The control unit 10 also includes a data acquisition unit 101, a classification model acquisition unit 102, a first learning unit (learning unit) 103, a ratio update unit 104, a first classification unit (classification unit) 105, a teacher data generation unit 106, a second learning unit 107, and a second classification unit 108. The storage unit 11 stores a first classification model (classification model) 111, teacher data 112, and a second classification model (classification model) 113.

[0015] The data acquisition unit 101 acquires various data used for machine learning. The data acquisition unit 101 also acquires data to be classified by the first classification model 111 and the second classification model 113. Here, the data to be classified is data that is known in advance to be biased in its classification.

[0016] "Data with a classification bias" refers to data for which the classification results are not uniform when each element contained in the data is classified into one of a plurality of classes. In other words, it is data with a significant difference in the ratio of elements classified into each class. For example, assume that a large number of data are classified into one of four classes. In this case, as the number of classified data increases, if the ratio of data classified into each class approaches 25% for all, the classification results are uniform, and it can be said that there is no bias in the classification of the data. On the other hand, as the number of classified data increases, if the data of elements classified into each class approaches, for example, 60%, 10%, 10%, and 20%, the classification results are non-uniform, and it can be said that there is a bias in the classification of the data.

[0017] The classification model acquisition unit 102 acquires the first classification model 111, which is a classification model to be adapted to a predetermined domain. As described above, the first classification model 111 is a classification model used for classifying data with a classification bias.

[0018] The first learning unit 103 updates the first classification model 111 using target data belonging to the predetermined domain in order to adapt the first classification model 111 obtained by the classification model acquisition unit 102 to the predetermined domain. Although details will be described later, target data not including correct labels is used for the update of the first classification model 111. In the update of the first classification model 111, the first learning unit 103 updates the first classification model 111 so that the number of target data classified into each class when a predetermined number of target data are classified into a plurality of classes by the first classification model 111 approaches a "predetermined ratio" obtained in advance according to the "bias in classification" in the target data. In this way, the first classification model 111 is updated by the first learning unit 103 and becomes a model adapted to the predetermined domain.

[0019] The ratio update unit 104 updates the "predetermined ratio" used for the subsequent updates of the first classification model 111 by the first learning unit 103 based on the bias in the classification result by the first classification model 111 updated by the first learning unit 103. The ratio update unit 104 is not an essential configuration in the information processing apparatus 1, but by providing the information processing apparatus 1 with the ratio update unit 104, when the bias in the classification of the data to be classified changes, it becomes possible to update the "predetermined ratio" following the change. And thereby, it becomes possible to update the first classification model 111 taking into account the above change. Therefore, when there is a possibility that the bias in the classification of the data to be classified changes, it is preferable to provide the ratio update unit 104.

[0020] The first classification unit 105 classifies data using the first classification model 111 updated by the first learning unit 103. The target data to be classified by the first classification unit 105 is data belonging to a predetermined domain to be adapted. And the classification result of the first classification unit 105 is used when generating teacher data 112 for adapting the second classification model 113 to the predetermined domain.

[0021] The teacher data generation unit 106 generates teacher data 112 for generating a second classification model 113 used for classifying data with a bias in classification by associating the classification result by the first classification unit 105 with the target data as a correct label. Note that re-learning and updating the learned second classification model 113 is also included in the scope of "generation" of the second classification model 113.

[0022] The second learning unit 107 generates the second classification model 113. More specifically, the second learning unit 107 generates the second classification model 113 using the teacher data 112 described above. The second classification model 113 generated using the teacher data 112 is adapted to a predetermined domain.

[0023] Thus, the second classification model 113 is a classification model generated by the second learning unit 107. The second classification model 113 is updated by supervised learning using the teacher data 112 by the second learning unit 107 and is adapted to a predetermined domain. The second classification model 113 only needs to be a classification model that can be generated and updated by supervised learning, and its classification algorithm is not particularly limited. For example, a neural network model or an FCM (Fuzzy C-Means) classifier may be used as the second classification model 113.

[0024] The second classification unit 108 classifies data using the second classification model 113. In particular, the second classification unit 108 can perform high-precision classification when the data to be classified has a bias similar to that of the target data described above. Also, the second classification unit 108 may also perform processing to output the classification result to the output unit 14 or the like.

[0025] As described above, the information processing apparatus 1 includes a classification model acquisition unit 102 that acquires a first classification model 111, which is a classification model to be adapted to a predetermined domain and is used for classifying data with a classification bias, and a first learning unit 103 that updates the first classification model 111 using target data that belongs to the above domain and does not include correct labels. Then, when the first learning unit 103 classifies a predetermined number of target data into a plurality of classes using the first classification model 111, the first classification model 111 is updated so that the number of target data classified into each class approaches a "predetermined ratio" obtained in advance according to the above bias.

[0026] According to the above configuration, since the first classification model 111 is updated using target data that does not include correct labels, it becomes possible to adapt the first classification model 111 to a predetermined domain without performing the operation of associating correct labels with the target data.

[0027] Also, according to the above configuration, when a predetermined number of target data are classified into each class, the first classification model 111 is updated so that the number of target data classified into each class approaches a "predetermined ratio" obtained in advance according to the "classification bias" in the data to be classified by the first classification model 111. Thereby, it is possible to update the first classification model 111 to adapt to a predetermined domain by utilizing the classification bias.

[0028] Therefore, according to the above configuration, it becomes possible to generate the first classification model 111 adapted to a predetermined domain without performing the operation of associating correct labels with the target data. Note that the "classification bias" can also be rephrased as the appearance ratio of each class in the classification result. That is, the first learning unit 103 updates the first classification model 111 so that the appearance ratio of each class in the classification result of the first classification model 111 after learning approaches the "predetermined ratio", and thereby, the first classification model 111 adapted to a predetermined domain is generated.

[0029] Also, as described above, the information processing apparatus 1 includes a first classification unit 105 that classifies target data belonging to a predetermined domain, and a teacher data generation unit 106 that generates teacher data 112 for generating a second classification model 113 used for classifying data with a bias in classification by associating the classification result by the first classification unit 105 with the target data as a correct label. Then, the first classification unit 105 classifies the target data by using a first classification model 111 generated by performing machine learning so that the number of target data classified into each class when a predetermined number of target data are classified into a plurality of classes approaches a predetermined ratio obtained in advance according to the above bias.

[0030] According to the above configuration, it is possible to generate teacher data 112 in which the classification result of the classification model is associated as a correct label from target data that does not include a correct label. Then, by performing machine learning using this teacher data 112, it is possible to generate a second classification model 113 adapted to the domain to which the target data belongs.

[0031] Also, according to the above configuration, when classifying a predetermined number of target data into a plurality of classes, the first classification model 111 that has been machine-learned so that the number of target data classified into each class approaches a predetermined ratio obtained in advance according to the bias in classification of the data to be classified is used. As a result, it is possible to classify target data with the above bias in classification with high accuracy, and by using this classification result, it is possible to generate teacher data 112 to which an appropriate correct label is associated. Furthermore, the second learning unit 107 can generate a second classification model 113 that can classify data with the above bias in classification with high accuracy by performing machine learning using this teacher data 112.

[0032] As described above, according to the above configuration, it is possible to adapt the second classification model 113 to a predetermined domain without performing the operation of associating a correct label with the target data.

[0033] (Generation of the First Classification Model and the Second Classification Model) The generation of the first classification model 111 and the second classification model 113 will be described with reference to FIG. 2. FIG. 2 is a diagram showing an example of the generation of the first classification model 111 and the second classification model 113.

[0034] In the following, an example will be described in which the first classification model 111 and the second classification model 113 are classification models for classifying images (hereinafter referred to as in-furnace images) that capture how waste burns in an incinerator in a waste incineration facility. Also, assume that the classes to be classified are the following four. Class 1: A state where there is not enough waste in the combustion stage Class 2: A state where there is appropriate garbage in the combustion stage and combustion is stable Class 3: A state where lumps of unburned waste are seen in the combustion stage or the afterburning stage and the flame is unstable Class 4: A state where it is difficult to classify into any of the above three states due to noise or the like The combustion state in the incinerator varies depending on the quality of the waste to be incinerated, etc. However, it has been empirically found that when the incinerator is operated continuously for a certain period, the appearance ratios of the above four classes converge to certain values. And it has also been empirically found that the converging values are not equal but are biased values. Therefore, it can be said that the in-furnace images are data with a recognized "classification bias". And according to this "classification bias", a "predetermined ratio" can be obtained in advance. For example, if the incinerator is operated for a period of one month to several months, and 10% of the in-furnace images taken at predetermined times during that period are Class 1, 60% are Class 2, 10% are Class 3, and 20% are Class 4, the "predetermined ratios" for Classes 1 to 4 may be set to 0.1, 0.6, 0.1, and 0.2.

[0035] By using the classification results of the above Classes 1 to 4 by the second classification model 113, the combustion state of the waste can be appropriately grasped. Also, by using this classification result, it is possible to automatically perform appropriate control on each device in the incineration facility.

[0036] Note that the data to be classified by the first classification model 111 and the second classification model 113 is arbitrary and is not limited to in-furnace images. Similarly, the classification classes of the first classification model 111 and the second classification model 113 are not particularly limited. The "in-furnace images" and "images" in the following description can be read as arbitrary data, and the classes in the following description can also be read as arbitrary classes.

[0037] In the example of FIG. 2, the first classification model 111 and the second classification model 113 are generated using a data set D1 including a plurality of in-furnace images and a data set D2 also including a plurality of in-furnace images (target data). Each in-furnace image included in the data set D1 may be, for example, an image taken at a predetermined incineration facility during a predetermined period. Each in-furnace image included in the data set D1 is associated with a correct label, that is, information indicating which of the above four classes the state inside the furnace shown in the in-furnace image corresponds to. That is, the data set D1 includes teacher data that can be used for supervised learning of a classification model for classifying in-furnace images. Note that the association of the correct labels in the data set D1 may be performed in advance by the user or the like.

[0038] Also, each in-furnace image included in the data set D2 is also an in-furnace image taken at a predetermined incineration facility during a predetermined period, similar to each in-furnace image included in the data set D1, but the shooting time is different from that of each in-furnace image included in the data set D1. For example, when each in-furnace image included in the data set D1 is an in-furnace image taken in a certain month, each in-furnace image included in the data set D2 may be an in-furnace image taken in the next month. It can be said that the data sets D1 and D2 belong to different domains because of the different shooting times. Another difference between the data set D2 and the data set D1 is that no correct label is associated with each in-furnace image included in the data set D2.

[0039] In the information processing apparatus 1, the data acquisition unit 101 acquires the data set D1. Then, the second learning unit 107 generates the second classification model 113 by machine learning using the data set D1. The second classification model 113 generated in this way is a classification model adapted to the data set D1, that is, a classification model capable of accurately classifying the in-furnace images in the same domain as each in-furnace image included in the data set D1. However, the classification accuracy of the second classification model 113 for the in-furnace images included in the data set D2 belonging to a domain different from the data set D1 is likely to be lower than the classification accuracy of the in-furnace images in the same domain as the data set D1.

[0040] Therefore, in order to enable the second learning unit 107 to accurately classify the data in the same domain as each in-furnace image included in the data set D2, the second classification model 113 is updated. In FIG. 2, the second classification model 113 adapted to the domain of the data set D2 is shown as the second classification model 113'.

[0041] In generating the second classification model 113', the data acquisition unit 101 acquires the data sets D1 and D2. Next, the first learning unit 103 generates the first classification model 111 using the acquired data sets D1 and D2. Although the details will be described in the item "Flow of the process of generating the first classification model 111" described later, the first classification model 111 generated using the data sets D1 and D2 is a classification model adapted to the domain of the data set D2, and it is possible to accurately classify the in-furnace images in the same domain as the data set D2.

[0042] Therefore, the first classification unit 105 classifies each in-furnace image included in the dataset D2 using the first classification model 111. Next, the teacher data generation unit 106 generates teacher data 112 by associating the classification result of the first classification unit 105 with each in-furnace image included in the dataset D2. Then, the second learning unit 107 retrains the second classification model 113 using the generated teacher data 112 to generate a second classification model 113' adapted to the domain of the dataset D2.

[0043] Here, after retraining using the dataset D2, as time further elapses, the tendency of the acquired in-furnace images may change. In this case, since the domain further changes, the classification accuracy of the second classification model 113' may decrease. According to the information processing apparatus 1, in such a case, a second classification model 113 adapted to the new domain can be generated in the same manner as in the example of FIG. 2. This will be described with reference to FIG. 3. FIG. 3 is a diagram showing another example of the update of the first classification model 111 and the second classification model 113.

[0044] In the example of FIG. 3, in addition to the above-described dataset D1, a dataset D3 is used. The dataset D3 includes in-furnace images taken at a predetermined incineration facility during a predetermined period, similar to the dataset D1. However, each in-furnace image included in the dataset D3 is different in terms of the shooting time from any of the in-furnace images included in the dataset D1 and the dataset D2 shown in FIG. 2. Also, similar to each image included in the dataset D2, no correct label is associated with each in-furnace image included in the dataset D3.

[0045] In the example of FIG. 3, the data acquisition unit 101 acquires data sets D1 and D3. Next, the first learning unit 103 generates a first classification model 111 using the acquired data sets D1 and D3. The first classification model 111 generated here is different from the first classification model 111 in FIG. 2 in that it is generated using data sets D1 and D3 and is thus adapted to the domain of data set D3. Therefore, in FIG. 3, the generated first classification model 111 is shown as the first classification model 111'.

[0046] Next, the first classification unit 105 classifies each in-furnace image included in the data set D3 using the first classification model 111'. Since the first classification model 111' is a classification model adapted to the domain of data set D3, it is possible to accurately classify data set D3.

[0047] Next, the teacher data generation unit 106 generates teacher data 112 by associating the classification results of the first classification unit 105 with each in-furnace image included in the data set D3. The teacher data 112 generated in the example of FIG. 3 is different from the teacher data 112 in FIG. 2 in that it is generated using data set D3. Therefore, in FIG. 3, the generated teacher data 112 is shown as teacher data 112'.

[0048] Then, the second learning unit 107 re-learns the second classification model 113' using the generated teacher data 112' and generates a second classification model 113'' adapted to the domain of data set D3. The second classification model 113'' generated in this way is a classification model adapted to data set D3, that is, a classification model capable of accurately classifying data in the same domain as data set D3.

[0049] In this way, by using the information processing apparatus 1, even when the tendency of the in-furnace images to be captured changes as the incineration facility itself and the waste to be incinerated change over time, it is possible to adapt the second classification model 113 to the change and maintain or improve the classification accuracy.

[0050] In addition, the information processing apparatus 1 can also be used when transferring the second classification model 113 generated for a certain incineration facility to another incineration facility using the in-furnace image taken at that incineration facility. In this case, the information processing apparatus 1 generates the first classification model 111 using the in-furnace image taken at a certain incineration facility and the in-furnace image taken at another incineration facility, and the in-furnace image taken at another incineration facility, the classification result of the first classification model 111 may be associated as the correct label to generate the teacher data 112. Then, the information processing apparatus 1 can generate the second classification model 113 adapted to other facilities by performing machine learning using this teacher data 112.

[0051] (Example of the first classification model 111) The first classification model 111 may have a configuration as shown in FIG. 4, for example. FIG. 4 is a diagram showing an example of the first classification model 111. The first classification model 111 shown in FIG. 4 includes a feature generator 1111, a first classifier 1112, and a second classifier 1113.

[0052] The feature generator 1111 generates feature information indicating the features of the input data input to the first classification model 111. The feature generator 1111 may have a configuration according to the data to be classified by the first classification model 111. For example, when classifying an in-furnace image, a feature generator 1111 that extracts feature amounts in the image may be used.

[0053] The first classifier 1112 classifies the input data using the feature information generated by the feature generator 1111. Also, the second classifier 1113 classifies the input data using the feature information generated by the feature generator 1111 in the same manner as the first classifier 1112. For example, when classifying the in-furnace image into the above-described four classes, the first classifier 1112 and the second classifier 1113 may be those that output numerical values indicating the probability that the input in-furnace image is classified into each class.

[0054] Note that although the second classifier 1113 and the first classifier 1112 share the same input and output (in other words, the explanatory variable and the objective variable), they are not the same classifier. Therefore, when the feature information generated by the feature generator 1111 is input to each of the first classifier 1112 and the second classifier 1113, differences may occur in the output classification results.

[0055] (Flow of the process for generating the first classification model 111) The flow of the process for generating the first classification model 111 will be described with reference to FIG. 5. FIG. 5 is a flowchart showing an example of the process for generating the first classification model 111. This flowchart includes each step of the method for generating the classification model according to the present embodiment.

[0056] In S11, the data acquisition unit 101 acquires a teacher data set and a target data set. The teacher data included in the teacher data set acquired here is not the teacher data 112 generated by the teacher data generation unit 106, but is obtained by associating the correct label with the input data whose correct label is known. For example, the data acquisition unit 101 may acquire data such as the data set D1 in FIG. 2 as the teacher data set. On the other hand, the target data set is a data set including data belonging to the domain to which the first classification model 111 is to be applied, and no correct label is associated with this data. For example, the data acquisition unit 101 may acquire the data set D2 in FIG. 2 or the data set D3 in FIG. 3 as the target data set.

[0057] Note that the method for acquiring the teacher data set and the target data set is not particularly limited. For example, the data acquisition unit 101 may acquire the teacher data set and the target data set stored in advance in the storage unit 11 or the like, or may acquire the teacher data set and the target data set via the communication unit 12 or the input unit 13. Also, the data acquisition unit 101 may acquire the teacher data set and the target data set by different methods respectively. Further, the timing at which the data acquisition unit 101 acquires these data sets does not necessarily have to be the same.

[0058] In S12 (acquisition step), the classification model acquisition unit 102 acquires a first classification model 111 including a feature generator 1111, a first classifier 1112, and a second classifier 1113. The first classification model 111 acquired in S12 may be one before learning or one that has already been learned. Also, the method for acquiring the first classification model 111 is not particularly limited. For example, the classification model acquisition unit 102 may acquire the first classification model 111 stored in the storage unit 11 or the like in advance, or may acquire the first classification model 111 via the communication unit 12 or the input unit 13.

[0059] In S13, the first learning unit 103 updates the first classification model 111 acquired in S12 using each piece of teacher data included in the teacher dataset acquired in S11. For example, the first learning unit 103 updates the feature generator 1111, the first classifier 1112, and the second classifier 1113 included in the first classification model 111 so that the cross-entropy error L when the teacher data is input to the first classification model 111 is minimized. This process can be represented by the following formula (1) with the feature generator 1111 as G, the first classifier 1112 as F 1 , the second classifier 1113 as F 2 , and the teacher dataset as {X s ,Y s}.

[0060]

Equation

[0061] Also, the cross-entropy error L in the above formula (1) is represented by the following formula (2) with the teacher data included in the teacher dataset as x s , the label of each piece of teacher data as y s , the classification target classes as 1 to K, and the probability that the classification result of x s is y as p.

[0062]

Equation

[0063] In S14 (update step), the first learning unit 103 updates the first classifier 1112 and the second classifier 1113 using each piece of target data included in the target data set acquired in S11. More specifically, the first learning unit 103 updates the first classifier 1112 and the second classifier 1113 so that the absolute value of the difference between the classification result of the first classifier 1112 and the classification result of the second classifier 1113 becomes larger. This process can be expressed by the following mathematical formula (3) with the target data set as {X t}.

[0064]

Number

[0065] In the above mathematical formula (3), L adv represents the absolute value of the difference between the classification result of the first classifier 1112 when classifying the target data and the classification result of the second classifier 1113. L adv is the probability p 1 (y|x) that the classification result is y when the element x is input to the first classifier 1112, the probability p 2 (y|x) that the classification result is y when the element x is input to the second classifier 1113, and with the target data x t included in the target data set {X t}, it is represented by the following mathematical formula (4).

[0066]

Number

[0067] The right side within the parentheses of mathematical formula (4) can be calculated by the following mathematical formula (5).

[0068]

Number

[0069] In mathematical formula (5), p1k is p for class k 1 , that is, the probability that the classification result of the first classifier 1112 is class k, and p 2k is p for class k 2 , that is, the probability that the classification result of the second classifier 1113 is class k. That is, when there are a total of K classes from 1 to K for the classification destination, p 1k and p 2k is to obtain the average value of the absolute value of the difference between them.

[0070] Also, the term L CB in the above formula (3) is a term for making the number of target data classified into each class approach a "predetermined ratio", and λ is a constant that serves as the weight for this term. L CB is represented by the following formula (6).

[0071]

Equation

[0072] The p k in formula (6) indicates the probability of being classified into class k, and is determined in advance according to the bias of the classification. For example, when the target data is an in-furnace image and this in-furnace image is classified into the above-mentioned 4 classes, it is known (or expected) that the probabilities of being classified into classes 1 to 4 are 60%, 10%, 10%, and 20% respectively. In this case, [p 1 , p 2 , p 3 , p 4 =[0.6, 0.1, 0.1, 0.2] can be set.

[0073] The value of L CB becomes 0 when the classification result of the target data is a predetermined ratio (for example, 60%, 10%, 10%, 20%), and becomes smaller (the absolute value becomes larger in the negative direction) as the classification result of the target data deviates from the predetermined ratio. Therefore, the first learning unit 103 is "-λ·L CBBy updating the first classifier 1112 and the second classifier 1113 so that the mathematical formula (3) including the term of "」" is minimized, when a predetermined number of target data are classified into a plurality of classes by the first classification model 111, the number of target data classified into each class approaches a "predetermined ratio" obtained in advance according to the bias of the classification of the target data, and thus the first classification model 111 can be updated.

[0074] In S15 (update step), the first learning unit 103 updates the feature generator 1111 using each target data included in the target data set acquired in S11. More specifically, the first learning unit 103 updates the feature generator 1111 so that the difference between the classification result of the first classifier 1112 and the classification result of the second classifier 1113 becomes small. This process can be represented by the following mathematical formula (7).

[0075]

Equation

[0076] According to the above mathematical formula (7), the feature generator 1111 is updated so that L adv (X t ) becomes minimum, that is, so that the "difference in classification results" when classifying the target data becomes minimum. Also, the above mathematical formula (7) includes the term of "-λ·L CB ". Therefore, by updating the feature generator 1111 so that the mathematical formula (7) is minimized, the first learning unit 103 can update the first classification model 111 so that the number of target data classified into each class approaches a "predetermined ratio" obtained in advance according to the bias of the classification of the target data when a predetermined number of target data are classified into a plurality of classes by the first classification model 111. In the example of FIG. 5, the process of S15 is performed after S14, but the process of S15 may be performed first and then the process of S14 may be performed.

[0077] In S16, the first learning unit 103 determines whether to finish updating the first classification model 111. If it is determined as NO in S16, the process returns to S13. If it is determined as YES in S16, the process proceeds to S17. The condition for finishing the update may be determined in advance. For example, the update may be finished when the processes of S13 to S15 are repeated a predetermined number of times. Note that the processes of S13 to S16 may be performed for each of a plurality of mini-batches obtained by dividing the teacher dataset and the target dataset. In this case, it is determined as YES in S16 when the update using all the mini-batches is finished.

[0078] In S17, the first learning unit 103 causes the storage unit 11 to store the updated first classification model 111. Thereby, the process of FIG. 5 ends.

[0079] As described above, the method for generating a classification model according to the present embodiment includes an acquisition step (S12) of acquiring a first classification model 111 that is a classification model to be adapted to a predetermined domain and is used for classifying data with a classification bias, and an update step (S14, S15) of updating the acquired first classification model 111 using target data belonging to the above domain and not including correct labels. In the update step, the first classification model 111 is updated so that the number of target data classified into each class when a predetermined number of target data are classified into a plurality of classes by the first classification model 111 approaches a "predetermined ratio" obtained in advance according to the classification bias of the target data. Thereby, it becomes possible to adapt the first classification model 111 to a predetermined domain without performing an operation of associating correct labels with the target data. Here, it is assumed that the data classified by the first classification model 111 all have a similar bias in the classification, including the target data. Therefore, the "classification bias of the target data" can also be referred to as the classification bias of the data classified by the first classification model 111.

[0080] Also, as described above, the classification model acquisition unit 102 may acquire, as the first classification model 111, a feature generator 1111 that generates feature information indicating the features of the target data, a first classifier 1112 that classifies the target data using the feature information, and a second classifier 1113 that classifies the target data using the feature information. Then, the first learning unit 103 performs a process of updating the first classifier 1112 and the second classifier 1113 so that the difference (specifically, the absolute value of the difference) in the classification results of the target data by the first classifier 1112 and the second classifier 1113 becomes large (S14 in FIG. 5), and a process of updating the feature generator 1111 so that the difference (specifically, the absolute value of the difference) in the classification results of the target data by the first classifier 1112 and the second classifier 1113 becomes small (S15 in FIG. 4).

[0081] According to the above configuration, the first classifier 1112 and the second classifier 1113 are updated so that the difference in the classification results of the target data by the first classifier 1112 and the second classifier 1113 becomes large. Thereby, the difference between the first classifier 1112 and the second classifier 1113 is clarified, and it becomes possible to detect target data (feature information) that is difficult to classify with the first classification model 111 before the update.

[0082] Also, according to the above configuration, the feature generator 1111 is updated so that the difference in the classification results of the target data by the first classifier 1112 and the second classifier 1113 becomes small. Thereby, the feature information generated from the target data (especially, the feature information generated from the target data that is difficult to classify with the first classification model 111 before the update) can be kept within a range where it can be correctly classified.

[0083] Therefore, according to the above configuration, the first classification model 111 can be adapted to the domain to which the target data belongs. Then, by using the updated first classification model 111 adapted to the domain to which the target data belongs, the target data can be classified accurately. Therefore, by labeling the classification result of the target data to the target data, it becomes possible to generate the teacher data 112 that can generate the second classification model 113 adapted to the domain.

[0084] Note that the method for updating the first classification model 111 by the first learning unit 103 is not limited to the above example. The update method updates the first classification model 111 using target data that does not include the correct label, and when a predetermined number of target data are classified into a plurality of classes by the first classification model 111, the number of target data classified into each class approaches a "predetermined ratio" obtained in advance according to the bias of the classification of the target data. For example, the first learning unit 103 may update the first classification model 111 using a well-known domain adaptation technique so that the number of target data classified into each class approaches the "predetermined ratio".

[0085] (Flow of the process for generating teacher data 112) The flow of the process for generating the teacher data 112 using the first classification model 111 generated as described above will be described with reference to FIG. 6. FIG. 6 is a flowchart showing an example of the process for generating the teacher data 112. This flowchart includes each step of the method for generating teacher data according to the present embodiment.

[0086] In S21, the data acquisition unit 101 acquires a target data set. Each target data included in the target data set is data that serves as the basis for the teacher data 112. The target data may be, for example, an in-furnace image. Note that no correct label is associated with the target data.

[0087] In S22 (classification step), the first classification unit 105 classifies each piece of target data included in the target data set acquired in S21 using the first classification model 111. For example, the first classification unit 105 may perform classification using the first classification model 111 including a feature generator 1111, a first classifier 1112, and a second classifier 1113 as shown in FIG. 4. In this case, the first classification unit 105 may use the classification result of either the first classifier 1112 or the second classifier 1113 as the final classification result, or may combine both classification results to generate the final classification result.

[0088] The first classification model 111 used in S22 is generated by performing machine learning so that when a predetermined number of target data belonging to a predetermined domain are classified into a plurality of classes, the number of target data classified into each class approaches a predetermined ratio obtained in advance according to the bias of the classification of the target data. For example, the first classification unit 105 may use the first classification model 111 generated and stored by the process of FIG. 5 for classification in S22.

[0089] In S23 (teacher data generation step), the teacher data generation unit 106 generates teacher data 112 using the classification result in S22. More specifically, the teacher data generation unit 106 generates teacher data 112 by associating the classification result in S22 with the target data acquired in S21 as a correct label. This process is performed for each piece of target data classified in S22.

[0090] In S24, the teacher data generation unit 106 stores the teacher data 112 generated in S23 in the storage unit 11.

[0091] In S25, the ratio update unit 104 determines whether to update the "predetermined ratio" referred to at the time of updating the first classification model 111 (S14, S15 in FIG. 5). If it is determined YES in S25, the process proceeds to S26, and if it is determined NO in S25, the process of FIG. 6 ends.

[0092] The determination condition in S25 may be determined in advance. For example, the ratio update unit 104 may calculate an index value indicating the degree of deviation between the ratio of the number of data classified into each class in S22 (hereinafter referred to as the "actual ratio") and the "predetermined ratio" referred to at the time of updating the most recent first classification model 111. Then, the ratio update unit 104 may determine whether or not to update based on whether the calculated index value is equal to or greater than a predetermined threshold value. As the index value indicating the degree of deviation, for example, a representative value (e.g., average value, maximum value, etc.) of the difference between the "actual ratio" and the "predetermined ratio" in each class may be used.

[0093] In S26, the ratio update unit 104 updates the "predetermined ratio" based on the bias of the classification result by the first classification model 111 updated by the first learning unit 103. For example, the ratio update unit 104 may use the above-described "actual ratio" as the new "predetermined ratio", or may use the average value or weighted average value of the "predetermined ratio" before update and the above-described "actual ratio" as the new "predetermined ratio". When the process of S26 ends, the process of FIG. 6 ends.

[0094] As described above, the method for generating teacher data according to the present embodiment includes a classification step (S22) of classifying target data of a predetermined number belonging to a predetermined domain into a plurality of classes by using a first classification model 111 generated by machine learning so that the number of target data classified into each class approaches a "predetermined ratio" obtained in advance according to the bias of the classification of the target data, and a teacher data generation step (S23) of generating teacher data 112 by associating the classification result in S22 with the target data as a correct label. Thereby, it becomes possible to generate teacher data 112 in which appropriate correct labels are associated with target data that does not include correct labels.

[0095] (Flow of processing for generating and updating the second classification model 113) The flow of processing for generating and updating the second classification model 113 will be described with reference to FIG. 7. FIG. 7 is a flowchart showing an example of the processing for generating and updating the second classification model 113.

[0096] In S31, the data acquisition unit 101 acquires a teacher data set. Each piece of teacher data included in this teacher data set is not generated by the teacher data generation unit 106, but is obtained by associating the correct label with input data whose correct label is known. For example, the teacher data set acquired in S11 of FIG. 5 may also be acquired in S31.

[0097] In S32, the second learning unit 107 generates a second classification model 113 using the teacher data set acquired in S31. Subsequently, in S33, the second learning unit 107 stores the second classification model 113 generated in S32 in the storage unit 11.

[0098] In S34, the data acquisition unit 101 determines whether new teacher data, that is, whether the teacher data 112 has been generated by the teacher data generation unit 106. If it is determined to be YES in S34, the process proceeds to the process of S35. On the other hand, if it is determined to be NO in S34, the process of S34 is performed again after a predetermined time has elapsed. Note that the process of S34 is not essential. For example, the processes after S35 may be performed when the teacher data generation unit 106 generates the teacher data 112, or the processes after S35 may be performed according to the user's input operation.

[0099] In S35, the data acquisition unit 101 acquires the teacher data 112. Note that there may be a plurality of pieces of teacher data 112 acquired in S35. Subsequently, in S36, the second learning unit 107 performs re-learning using the teacher data 112 acquired in S35 and updates the second classification model 113 stored in S33. Then, in S37, the second learning unit 107 stores the second classification model 113 updated in S36 in the storage unit 11, and thereby the process of FIG. 7 ends.

[0100] Note that instead of performing re-learning in S36, the second learning unit 107 may generate a new second classification model 113 using the teacher data 112. In this case, the second learning unit 107 may generate the second classification model 113 using both the teacher data acquired in S31 and the teacher data 112 acquired in S35.

[0101] 〔Embodiment 2〕 Other embodiments of the present invention will be described below. For convenience of explanation, members having the same functions as those described in the above embodiments are denoted by the same reference numerals, and their descriptions will not be repeated.

[0102] (Configuration of the information processing apparatus 1A) FIG. 8 is a block diagram showing an example of the main configuration of the information processing apparatus 1A according to the present embodiment. As shown in the figure, the information processing apparatus 1A includes a control unit 10A that comprehensively controls each part of the information processing apparatus 1A, and a storage unit 11A that stores various data used by the information processing apparatus 1A. Further, the control unit 10A includes a data acquisition unit 101A, a classification model acquisition unit 102A, a learning unit 103A, and a classification unit 105A. Also, a classification model 111A is stored in the storage unit 11A. Note that the information processing apparatus 1A may include a ratio update unit 104.

[0103] The data acquisition unit 101A acquires various data used for machine learning, similarly to the data acquisition unit 101 in the information processing apparatus 1. Further, the data acquisition unit 101A acquires data to be classified by the classification model 111A.

[0104] The classification model acquisition unit 102A acquires a classification model 111A that is a classification model to be adapted to a predetermined domain and is used for classifying data with a bias in classification, similarly to the classification model acquisition unit 102 in the information processing apparatus 1.

[0105] Similar to the first learning unit 103 in the information processing apparatus 1, the learning unit 103A updates the classification model 111A acquired by the classification model acquisition unit 102A using target data belonging to a predetermined domain in order to adapt the classification model 111A to the predetermined domain.

[0106] Similar to the first classification unit 105 in the information processing apparatus 1, the classification unit 105A classifies data using the classification model 111A updated by the learning unit 103A.

[0107] The classification model 111A is a classification model to be adapted to a predetermined domain, similar to the first classification model 111 in the information processing apparatus 1, and is a classification model used for classifying data with a bias in classification. As described above, the classification model 111A is updated by the learning unit 103A to become a model adapted to a predetermined domain. The classification model 111A may have a configuration including a feature generator, a first classifier, and a second classifier, for example, like the first classification model 111 shown in FIG. 4.

[0108] (Generation of Classification Model 111A) The generation of the classification model 111A will be described with reference to FIG. 9. FIG. 9 is a diagram showing an example of the generation of the classification model 111A. Here, similar to the description of FIGS. 2 and 3, an example in which the classification model 111A is a classification model for classifying in-furnace images will be described.

[0109] In the example of FIG. 9, first, a dataset D1 is acquired, and then datasets D2 and D3 are sequentially acquired over time. A correct label is associated with each in-furnace image included in the dataset D1, but no correct label is associated with each in-furnace image included in the datasets D2 and D3.

[0110] The learning unit 103A generates a classification model 111A using the dataset D1 as a teacher dataset. Since the classification model 111A is generated by machine learning using the dataset D1, it becomes a classification model adapted to the domain of the dataset D1. Therefore, as long as there is no change in the tendency of the in-furnace images to be captured, in other words, as long as the domain of the in-furnace images does not change, the classification unit 105A can appropriately classify newly acquired in-furnace images using the classification model 111A.

[0111] However, it is common for the tendency of the in-furnace images to be captured to change over time, and it can be said that the domain of the in-furnace images changes after a certain period of time. Therefore, the data acquisition unit 101A starts collecting new in-furnace images when a predetermined time has elapsed since each in-furnace image included in the dataset D1 was captured. The dataset D2 shown in FIG. 9 is the in-furnace image collected in this way. Similarly, the dataset D3 shown in FIG. 9 is the in-furnace image collected at a time point later than the dataset D2. Therefore, it can be said that the datasets D1, D2, and D3 belong to different domains. Note that, similar to the first embodiment, it is not necessary to associate a correct label with each in-furnace image included in the datasets D2 and D3.

[0112] When the dataset D2 is acquired, the learning unit 103A updates the classification model 111A using the acquired dataset D2 (including the target data). Note that the dataset D1 may also be used for updating the classification model 111A. In FIG. 9, the updated classification model 111A is shown as the classification model 111A'. The classification model 111A' is a classification model adapted to the domain of the dataset D2. Note that the same method as in the first embodiment can be used as the update method.

[0113] Similarly, when the dataset D3 is acquired, the learning unit 103A updates the classification model 111A’ using the acquired dataset D3 (target data). Note that the dataset D1 may also be used for updating the classification model 111A’. In FIG. 9, the updated classification model 111A is shown as the classification model 111A”. The classification model 111A” is a classification model adapted to the domain of the dataset D3. Note that the learning unit 103A may update the classification model 111A instead of the classification model 111A’ to generate the classification model 111A”.

[0114] In this way, since the information processing apparatus 1A updates the classification model 111A using the dataset obtained over time, it is possible to maintain or improve the classification accuracy even when the domain of the in-furnace image changes with time measurement.

[0115] (Flow of the process for generating the classification model 111A) The flow of the process for generating the classification model 111A will be described based on FIG. 10. FIG. 10 is a flowchart showing an example of the process for generating the classification model 111A. This flowchart includes each step of the method for generating the classification model according to the present embodiment.

[0116] In S51, the data acquisition unit 101A acquires a teacher dataset for generating the classification model 111A. The teacher data acquired here is obtained by associating the correct label with the input data whose correct label is known. For example, the data acquisition unit 101A may acquire the dataset D1 in FIG. 9 as the teacher dataset.

[0117] In S52, the learning unit 103A generates the classification model 111A using the teacher data acquired in S51. A general supervised learning method can be applied to the generation of the classification model 111A in S52. Subsequently, in S53, the learning unit 103A stores the classification model 111A generated in S52 in the storage unit 11A.

[0118] In S54, the data acquisition unit 101A determines whether the classification model 111A needs to be updated. The determination condition may be set in advance. For example, the data acquisition unit 101A may determine that an update is required when a predetermined period has elapsed since the previous update. If it is determined as YES in S54, the process proceeds to S55. If it is determined as NO in S54, the process of S54 is performed again after a predetermined time has elapsed. Note that the process of S54 is not essential. For example, the processes after S55 may be performed according to a user input operation.

[0119] In S55, the data acquisition unit 101A acquires a target data set. Each target data included in the target data set is data belonging to the domain to be adapted and does not include a correct label. For example, the data acquisition unit 101A may acquire the data set D2 or D3 in FIG. 9 as the target data set.

[0120] In S56 (acquisition step), the classification model acquisition unit 102A acquires the classification model 111A stored in the storage unit 11A in S53.

[0121] In S57 (update step), the learning unit 103A updates the classification model 111A acquired in S56 using the target data acquired in S55. The update method may be any method that can generate a classification model 111A adapted to the domain of the target data using target data that does not include a correct label. For example, when updating the classification model 111A including a feature generator, a first classifier, and a second classifier as shown in the first classification model 111 in FIG. 4, the learning unit 103A may update the classification model 111A by the processes of S13 to S16 in FIG. 5.

[0122] In S58, the learning unit 103A stores the updated classification model 111A in the storage unit 11A. Thereby, the process of FIG. 10 ends.

[0123] 〔Modification example〕 The execution entity of each process described in each of the above embodiments is arbitrary and is not limited to the above examples. That is, the functions similar to those of information processing apparatus 1 or 1A can be realized by a plurality of information processing apparatuses capable of communicating with each other. For example, the processes shown in FIGS. 5 to 7 and FIG. 10 may be executed by a plurality of information processing apparatuses in a shared manner. Also, for example, the processes in FIG. 5, the processes in FIG. 6, and the processes in FIG. 7 may be executed by different information processing apparatuses respectively. That is, the apparatus that updates the first classification model 111 (which can also be called the first learning apparatus), the apparatus that generates the teacher data 112 (which can also be called the teacher data generation apparatus), and the apparatus that generates the second classification model 113 (which can also be called the second learning apparatus) may each be an individual apparatus.

[0124] 〔Example of Realization by Software〕 The functions of information processing apparatuses 1 and 1A are programs for causing a computer to function as information processing apparatus 1 or 1A, and can be realized by programs (learning programs / teacher data generation programs) for causing a computer to function as each control block of information processing apparatus 1 or 1A (especially each part included in control units 10 and 10A).

[0125] In this case, information processing apparatus 1 or 1A includes, as hardware for executing the above program, a computer having at least one control device (for example, a processor) and at least one storage device (for example, a memory). By executing the above program with this control device and storage device, each function described in each of the above embodiments is realized.

[0126] The above program may be recorded on one or more computer-readable recording media, not temporarily. This recording medium may or may not be provided in information processing apparatus 1 or 1A. In the latter case, the above program may be supplied to the above apparatus via any wired or wireless transmission medium.

[0127] In addition, part or all of the functions of each of the above control blocks can also be realized by a logic circuit. For example, an integrated circuit in which a logic circuit functioning as each of the above control blocks is formed is also included in the scope of the present invention. In addition to this, for example, it is also possible to realize the functions of each of the above control blocks by a quantum computer.

[0128] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope shown in the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention.

Explanation of Signs

[0129] 1 Information processing apparatus 102 Classification model acquisition unit 103 First learning unit (learning unit) 104 Ratio update unit 105 First classification unit (classification unit) 106 Teacher data generation unit 111 First classification model (classification model) 112 Teacher data 113 Second classification model (classification model) 1111 Feature generator 1112 First classifier 1113 Second classifier 1A Information processing apparatus 102A Classification model acquisition unit 103A Learning unit 111A Classification model

Claims

1. An information processing apparatus, comprising: a classification model acquisition unit that acquires a classification model that is a target for adaptation to a predetermined domain and is used for classifying data with bias in classification; a learning unit that updates the classification model using target data that belongs to the domain and does not include correct labels; wherein the learning unit updates the classification model such that the number of pieces of target data classified into each class when a predetermined number of pieces of target data are classified into a plurality of classes by the classification model approaches a predetermined ratio obtained in advance according to the bias.

2. The classification model acquisition unit acquires, as the classification model, a feature generator that generates feature information indicating features of the target data, a first classifier that classifies the target data using the feature information, and a second classifier that classifies the target data using the feature information; The learning unit performs a process of updating the first classifier and the second classifier so that the difference in the classification results of the target data by the first classifier and the second classifier becomes large, and performs a process of updating the feature generator so that the difference in the classification results of the target data by the first classifier and the second classifier becomes small, the information processing apparatus according to claim 1.

3. The information processing apparatus according to claim 1 or 2, further comprising a ratio update unit that updates the predetermined ratio based on the bias of the classification result by the classification model updated by the learning unit.

4. a classification unit that classifies target data belonging to a predetermined domain; a teacher data generation unit that associates the result of the classification with the target data as a correct label and generates teacher data for generating a classification model used for classifying data with bias in classification; wherein the classification unit classifies the target data using a classification model generated by performing machine learning such that the number of pieces of target data classified into each class when a predetermined number of pieces of target data are classified into a plurality of classes approaches a predetermined ratio obtained in advance according to the bias.

5. A method for generating a classification model, executed by at least one information processing apparatus, the method comprising: an acquisition step of acquiring a classification model that is a target for adaptation to a predetermined domain and is used for classifying data with bias in classification; An update step of updating the classification model using target data that belongs to the domain and does not include correct labels, In the update step, the classification model is updated so that the number of pieces of target data classified into each class when a predetermined number of pieces of target data are classified into a plurality of classes by the classification model approaches a predetermined ratio obtained in advance according to the bias. A method for generating a classification model.

6. A method for generating teacher data used for generating a classification model used for classifying data with a bias in classification, which is executed by at least one information processing device, A classification step of classifying the target data by a classification model generated by machine learning so that the number of pieces of target data classified into each class when a predetermined number of pieces of target data belonging to a predetermined domain are classified into a plurality of classes approaches a predetermined ratio obtained in advance according to the bias, A teacher data generation step of generating teacher data by associating the result of the classification with the target data as a correct label. A method for generating teacher data.

7. A learning program for causing a computer to function as the information processing device according to claim 1, the learning program for causing a computer to function as the classification model acquisition unit and the learning unit.

8. A teacher data generation program for causing a computer to function as the information processing device according to claim 4, the teacher data generation program for causing a computer to function as the classification unit and the teacher data generation unit.

Images