JP2023011312A - Machine learning device, machine learning method, and machine learning program - Google Patents

Abstract

To provide a machine learning technique capable of improving classification accuracy and memory use efficiency in incremental learning that combines an initial session for performing clustering of a base class with an additional session for performing clustering of a new class.SOLUTION: A feature extraction unit 10 extracts a feature vector of an input sample using a learned model. A clustering unit 40 performs clustering of the input sample based on a feature vector of the input sample and a class to which the input sample belongs to generate a per-class centroid. The clustering unit 40 discriminates a parameter used in clustering between an initial session for clustering a sample of a training data set of a base class used in generating a learned model and an additional session for clustering a sample of a training data set of a new class to be newly added.SELECTED DRAWING: Figure 2

Description

The present invention relates to machine learning technology.

Humans can learn new knowledge through long-term experience, and can maintain old knowledge without forgetting it. On the other hand, the knowledge of the convolutional neural network (CNN) depends on the dataset used for training, and in order to adapt to changes in the data distribution, the CNN parameters must be retrained for the entire dataset. Is required. As the CNN learns about new tasks, its estimation accuracy for old tasks decreases. In this way, continuous learning in a CNN inevitably causes catastrophic forgetting, in which learning results of old tasks are forgotten during learning of new tasks.

As a method for avoiding fatal forgetting, class incremental learning has been proposed, in which a new class is additionally learned to a trained model, and a new class can be classified together with the existing class.

Patent Document 1 proposes a technique called CBCL (Centroid-Based Concept Learning) as a cognitive inspiration model for incremental learning. In CBCL, features extracted from data samples are clustered and stored in the form of centroids, and label prediction is performed by weighted voting based on neighboring centroids.

A. Ayub and A. R. Wagner, "Cognitively-Inspired Model for Incremental Learning Using a Few Examples," 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020, pp. 897-906.

In CBCL, the pretrained model is used to cluster the samples of the new class training data set, but not the samples of the base class training data set used to generate the pretrained model. When extending CBCL to incremental learning that combines the initial session for clustering the base class and the additional session for clustering the new class, if the parameters used for clustering are common in the initial session and the additional session, the centroid of the base class There is likely to be a bias in the numbers and the number of centroids in the new class. As a result, it is expected that the centroid-based voting will not be conducted fairly between the base class and the new class, and the classification accuracy will deteriorate. In addition, since the optimal number of centroids is considered to differ between the base class and the new class, it is expected that memory usage required for storing the centroids will also be wasted.

The present invention has been made in view of this situation, and its object is to improve classification accuracy and memory usage efficiency in incremental learning that combines an initial session for clustering the base class and an additional session for clustering the new class. The purpose is to provide a machine learning technology that can realize

In order to solve the above problems, a machine learning device according to one aspect of the present invention includes a feature extraction unit that extracts a feature vector of an input sample using a trained model, and a feature vector of the input sample and the input sample to which the feature vector belongs. a clustering unit that clusters the input samples based on the classes to generate centroids for each class. The clustering unit includes an initial session for clustering the base class training data set samples used to generate the trained model, and an additional session for clustering the newly added new class training data set samples, Different parameters are used for clustering.

Another aspect of the invention is a machine learning method. This method includes a feature extraction step of extracting a feature vector of an input sample using a trained model, and clustering the input sample based on the feature vector of the input sample and the class to which the input sample belongs to cluster the cents of each class. and a clustering step to generate the roids. The clustering step includes an initial session for clustering the base class training data set samples used in generating the trained model and an additional session for clustering the newly added new class training data set samples, Different parameters are used for clustering.

Any combination of the above constituent elements, and any conversion of expressions of the present invention into methods, devices, systems, recording media, computer programs, etc. are also effective as embodiments of the present invention.

According to the present invention, classification accuracy can be improved and memory usage efficiency can be improved in incremental learning in which an initial session for clustering a base class and an additional session for clustering a new class are combined.

FIG. 4 is a diagram for explaining the flow of incremental learning processing according to the first embodiment; 1 is a configuration diagram of a machine learning device according to a first embodiment; FIG. 3(a) and 3(b) are flow charts for explaining the machine learning procedure by the machine learning apparatus of FIG. FIG. 11 is a configuration diagram of an inference device according to a second embodiment; 5 is a flowchart for explaining an inference procedure by the inference device of FIG. 4;

(First embodiment)
CBCL uses pretrained models as feature extractors without updating them. CBCL is not premised on classifying pre-trained classes (called "base classes") used to generate pre-trained models, and new classes unrelated to pre-learning ("new classes" ) is assumed to be continuously learned.

On the other hand, in general, many methods of incremental learning perform learning so as to enable additional new class classification while maintaining the classification ability of the base class. Algorithms have been devised that take into account the prevention of catastrophic forgetting so that not only can classification of new classes become possible, but also the classification accuracy of base classes does not deteriorate significantly.

CBCL is not an algorithm premised on base class clustering, but in the first embodiment, it is premised on extending CBCL to clustering of both base classes and new classes like general incremental learning. .

FIG. 1 is a diagram for explaining the flow of processing when CBCL is extended to incremental learning in which base class clustering and new class clustering are combined.

Pre-trained models generated using the base class training data set 300 are pre-selected as feature extractors (S10). The pre-trained model is not updated in subsequent processing and is fixed.

A series of processes for applying the CBCL algorithm to the base class is called an "initial session". In the initial session, a base class training data set 300 is input and a centroid is generated for each class for the base class (S20). The data sample input in the initial session is preferably all or part of the training data used in pre-learning, or data in the same class as the training data used in pre-learning and having similar features. The number of base classes is affected by the number of classes that can be classified by pre-learning.

A series of processes for applying the CBCL algorithm to a new class is called an "additional session". After the initial session, additional sessions input new class training data sets 310, 320, 330 for each session and generate class-by-class centroids for the new classes (S21, S22, S23). The number of new classes is not limited. It is assumed that the additional sessions are continuously repeated in session units, and the additional session S1 (S21) for the new class C1, the additional session S2 (S22) for the new class C2, the additional session S3 (S23) for the new class C3, and so on. and continues. A centroid is generated for each class for each new class in each additional session. At this time, each session inherits the centroid for each class of the previous session. The new classes to be dealt with in the later sessions are mainly the classes that have not already been introduced, but it is possible that the classes that have already been introduced will be dealt with again.

FIG. 2 is a configuration diagram of the machine learning device 100 according to the first embodiment. The machine learning device 100 includes a feature extraction unit 10 , a trained model storage unit 20 , a centroid generation unit 30 , a centroid storage unit 50 and a label prediction unit 60 . The centroid generation unit 30 includes a clustering unit 40 , and the label prediction unit 60 includes a neighborhood centroid selection unit 70 and a weighted voting unit 80 .

The machine learning device 100 receives input of training data, learns clustering by the CBCL technique, generates centroids for each class, and performs label prediction for unknown data samples based on the centroids. If a series of learning processes for a certain training data set is regarded as a session, CBCL assumes that learning is continuously repeated for each session. The number of classes in the training dataset in each session is unlimited. The processing contents of the session to which the CBCL algorithm is applied will be described with reference to FIG. 1, but the processing contents are the same in the initial session and each additional session.

In each session, when a training data set to be learned is input to the machine learning device 100, the feature extraction unit 10 reads out a pre-selected pre-trained model from the learned model storage unit 20, and extracts the pre-trained model. It is used as a feature extractor to extract feature vectors of the input samples of the training data set. An example of a sample is an image, but a sample is not necessarily limited to an image.

In the learning phase, the feature extraction unit 10 supplies the extracted feature vectors of the input samples to the centroid generation unit 30 . The centroid generation unit 30 clusters the input samples based on the feature vector of the input sample and the label of the class to which the input sample belongs, generates a centroid for each class, and stores the generated centroid for each class in the centroid storage unit. Store in 50.

The centroid generation unit 30 includes a clustering unit 40, and generates a centroid for each class by repeating machine learning called Agg-Var clustering, for example. The number of centroids for each class is one or more and is not constant for all classes.

The clustering unit 40 calculates the distance between the feature vector of the input sample and each centroid of the class to which the input sample belongs, and finds the closest neighbor to the feature vector of the input sample among a plurality of centroids of the class to which the input sample belongs. Select a centroid. The clustering unit 40 compares the distance dmin between the input sample and the nearest centroid with a threshold value D to integrate the feature vector of the input sample into the nearest centroid or classify the feature vector of the input sample into the class to which the input sample belongs. Decide whether to add a new centroid for Specifically, when the distance dmin between the input sample and the nearest centroid is less than a threshold D (dmin<D), the clustering unit 40 integrates the feature vectors of the input sample with the nearest centroid, and the distance dmin is the threshold If D or more (dmin≧D), the feature vector of the input sample is newly added as the centroid of the class to which the input sample belongs.

Applying Agg-Var clustering to all of the training data sets for each session establishes the centroids for all classes in that session. Since the centroid of the previous session is inherited, the centroid is updated in the latter session while including the centroid of the class learned up to the previous session. As described above, when dealing with an existing class in a subsequent session, the Agg-Var clustering process is performed with the centroid already generated for the existing class as an initial state.

In the inference phase, the feature extraction unit 10 supplies feature vectors extracted from unknown samples to the label prediction unit 60 . The label prediction unit 60 predicts the class label of each sample based on the determined centroids stored in the centroid storage unit 50 .

In the label prediction unit 60, the neighborhood centroid selection unit 70 selects n neighborhood centroids for the feature vector of the unknown sample, considering the distances from all centroids of each class.

The weighted voting unit 80 assigns a weight according to the distance between the feature vector of the unknown sample and the n-neighboring centroids to the candidate class to which each of the n centroids selected by the neighborhood centroid selection unit 70 belongs. Weighted voting is performed by calculating the prediction score Pred(y) for each candidate class y with . For example, the reciprocal of the distance is used as the weight according to the distance. That is, the shorter the distance, the greater the weight given to the vote.

The weighted voting unit 80 predicts the label assigned to the candidate class having the maximum prediction score Pred(y) as the class label of the unknown sample.

The method described in Patent Document 1 can be used for CBCL processing such as Agg-Var clustering and prediction score calculation.

In conventional CBCL, the threshold value D, which is a parameter of Agg-Var clustering, is one in one learning and is a fixed value that is not particularly controlled. Here, let us consider a problem when processing each session described with reference to FIGS. 1 and 2 with the threshold value D fixed according to the conventional CBCL.

Training data sets are usually selected so that there is no large bias in the number of samples per class or feature distribution in each set. However, when comparing the base class dataset and the new class dataset, the number of samples and feature distributions for each class are likely to be different.

In the CBCL algorithm, it is desirable to use a model with excellent feature extraction function as a trained model, so the trained model should be based on big data, and the base class training data is For example, it is assumed that there are several hundred samples or more. On the other hand, the training data for the new class may be big data, but preferably small data, that is, the number of samples is small. In particular, as described in Patent Document 1, ideally, the Few-Shot format, ie, training data of, for example, several samples per class. Learning using big data has practical issues such as the time and effort required to prepare a large number of samples in advance and the long learning time. Incremental learning using is preferred.

Therefore, in FIG. 1, if the threshold D, which is a parameter of Agg-Var clustering, is common in both the initial session and the additional session, it is thought that there will be a bias in the number of centroids between the base class and the new class. be done. In general, it is expected that the number of centroids will be large if the base class is big data, and the number of centroids will be small if the new class is small data. In that case, it is expected that in the label prediction process, the centroid-based vote will be dominant over the base class, and the vote will not be impartially performed between the base class and the new class, degrading the classification accuracy. In addition, although many base class centroids will be saved, it is expected that memory usage efficiency will decrease because many unnecessary centroids will be included compared to new class centroids.

Therefore, in the first embodiment, the threshold D, which is a parameter of Agg-Var clustering, is set in the initial session for clustering the base class training data set and the additional session for clustering the new class training data set. Set a suitable value. Regarding the threshold Df for the initial session and the threshold Di for the additional session, the threshold Df is set higher than the threshold Di, and control is performed so that the base class centroid is reduced as much as possible. In other words, the threshold Di is set smaller than the threshold Df, and control is performed to increase the number of new class centroids as much as possible. As a result, the fairness of voting based on the centroid can be improved and the classification accuracy can be improved compared to the case where the threshold D is shared between the base class and the new class. In addition, since the number of centroids in the base class is reduced, unnecessary centroids do not need to be stored, so memory usage can be made more efficient.

It is desirable that the initial session threshold Df and the additional session threshold Di have similar distributions of distances between centroids in a class. For example, the threshold Df and the threshold Di are set based on the probability distribution of the feature data based on the feature data extracted using the original data of the training data or the pre-trained model. Assuming that the probability distribution follows a normal distribution, the larger the average of the squares of the variances σ of the training data of the base class for each class, the greater the difference between the values of the threshold Df and the threshold Di. set to be As a result, the number of centroids per class of the base class can be reduced as much as possible.

3(a) and 3(b) are flowcharts for explaining the machine learning procedure by the machine learning device 100 of FIG.

FIG. 3(a) shows the flow of clustering processing in the learning stage. The threshold value Df of the initial session is set to the threshold value D, which is a parameter used for clustering (S30). An initial session clustering process is executed (S32). The threshold Di of the additional session is set to the threshold D (S34). Clustering processing of the additional session is executed (S36). If the next additional session is to be clustered (Y of S38), return to step S36; otherwise (N of S38), end.

FIG. 3(b) shows detailed procedures of the initial session clustering process in step S32 and the additional session clustering process in step S36 of FIG. 3(a). The clustering process is the same for initial and additional sessions. The initial session clustering process is input with samples of the base class training data set, and the additional session clustering process is input with samples of the new class training data set.

The feature extraction unit 10 extracts the feature vector of the input sample using the trained model (S40).

The clustering unit 40 selects the nearest neighbor centroid closest to the feature vector of the input sample from among the plurality of centroids of the class to which the input sample belongs (S42). The clustering unit 40 compares the distance dmin between the feature vector of the input sample and the nearest centroid with a threshold value D (S44).

If dmin<D (Y in S46), the clustering unit 40 integrates the feature vectors of the input samples into the nearest centroid (S48). If dmin≧D (N of S46), the feature vector of the input sample is newly added as a centroid of the class to which it belongs (S50).

(Second embodiment)
FIG. 4 is a configuration diagram of an inference device 200 according to the second embodiment. Inference device 200 includes feature extraction unit 10 , trained model storage unit 20 , centroid generation unit 30 , centroid storage unit 50 , and label prediction unit 60 . The centroid generation unit 30 includes a clustering unit 40 , and the label prediction unit 60 includes a neighborhood centroid selection unit 70 and a weighted voting unit 80 .

The machine learning apparatus 100 of the first embodiment and the inference apparatus 200 of the second embodiment differ in the configuration and operation of the clustering unit 40 and the weighted voting unit 80, and are otherwise common in configuration and operation.

The clustering unit 40 of the machine learning device 100 of the first embodiment differentiates the threshold value D, which is a parameter used in clustering, between the initial session and the additional session. The parameters used for clustering may be common for all additional sessions without clustering the base class as in conventional CBCL. Alternatively, when the inference apparatus 200 of the second embodiment also performs clustering of the initial session as in the first embodiment, the clustering unit 40 of the inference apparatus 200 also sets the threshold value D, which is a parameter used for clustering, to It may be different for initial sessions and additional sessions.

The configuration and operation of the weighted voting unit 80 in the label prediction unit 60 of the inference device 200 according to the second embodiment will be described below.

The weighted voting unit 80 weights the candidate classes to which each of the n centroids selected by the neighborhood centroid selection unit 70 belongs according to the distance between the feature vector of the unknown sample and the centroid. A prediction score Pred(y) for each candidate class y is calculated, and the prediction score Pred(y) is multiplied by a correction coefficient A to calculate a corrected prediction score Pred(y)′, and the corrected prediction score Pred(y)′ is calculated. Predict the label assigned to the class with the maximum value as the label of the unknown sample.

Here, in the conventional CBCL, this correction factor A is the reciprocal of the number _Ny of samples of the training data of the class. That is, A=1/N _y and Pred(y)′=(1/N _y )Pred(y).

This correction coefficient A = 1/N _y is likely to increase the number of centroids for classes with a large number of samples. It is based on the premise that there is and is unfair.

Here, the training data set is usually selected so that the number of samples for each class and the feature distribution are not greatly biased in each set. However, as an ideal application, it is desirable to repeat sessions semipermanently, so the number of samples for each class and feature distribution are not always uniform in all sessions. In addition, conventional CBCL is not premised on classifying the base class and assumes continuous learning of new classes, but there are applications that extend CBCL to classify both base classes and new classes. I can think of many. In that case, the base class is big data (for example, several hundred samples or more per class), and the new class is small data, ideally in Few-Shot format (for example, about several samples per class). many.

If the number of samples and the feature distribution are the same for all classes in all sessions, the correction coefficient A can be A=1/N _y as in the conventional CBCL. Otherwise, if the correction coefficient A is 1/N _y , the prediction score Pred(y)' of the candidate class with a larger number of samples will be smaller, making it more difficult for that candidate class to be selected as a label. A candidate class with a small number of samples has a large prediction score Pred(y)′, and the candidate class is more likely to be selected as a label. Therefore, it is expected that the number of samples for each class will not give fair voting, and that the classification accuracy will deteriorate.

Therefore, in the second embodiment, the correction coefficient A by which Pred(y) is multiplied reflects the number of centroids for each class. Assuming that the number of centroids for each class is N ^* _y , and the correction coefficient A is, for example, 1/N ^* _y , which is the reciprocal of the number of centroids, the influence that each centroid has on the prediction of the class label, that is, the energy is become the same. As a result, even if the number of samples and feature distribution are not the same for each class, the fairness of voting based on the centroid can be improved and the deterioration of classification accuracy can be prevented.

Thus, in the second embodiment, the weighted voting unit 80 multiplies the prediction score Pred(y) for each class y by the correction coefficient A=1/N ^* _y to obtain the corrected prediction score Pred(y) ' is calculated, and the class having the maximum corrected prediction score Pred(y)' is predicted as the class label of the unknown sample.

FIG. 5 is a flowchart for explaining the inference procedure by the inference apparatus 200 of FIG.

The feature extraction unit 10 extracts the feature vector of the unknown sample using the trained model (S70).

The neighborhood centroid selection unit 70 selects n neighborhood centroids for the feature vector of the unknown sample (S72).

The weighted voting unit 80 calculates the prediction score of the candidate class to which the n-neighboring centroids belong by adding weights according to the distances between the feature vector of the unknown sample and the n-neighboring centroids (S74). The weighted voting unit 80 corrects the predicted score of the candidate class with a correction coefficient according to the number of centroids of the candidate class (S76). The weighted voting unit 80 predicts that the class with the maximum corrected prediction score is the class label (S78).

The various processes of the machine learning device 100 and the reasoning device 200 described above can of course be realized as a device using hardware such as a CPU and memory, and can also be implemented using a ROM (read only memory) or flash memory. It can also be realized by firmware stored in a memory or the like, or software such as a computer. The firmware program or software program may be recorded on a computer-readable recording medium and provided, transmitted to or received from a server via a wired or wireless network, or transmitted or received as data broadcasting of terrestrial or satellite digital broadcasting. is also possible.

As described above, according to the machine learning device 100 of the first embodiment, when CBCL is extended to incremental learning in which an initial session for clustering the base class and an additional session for clustering the new class are combined, By using different parameters for clustering between the base class and the new class, it is possible to improve the classification accuracy of the base class and the new class, reduce the number of wasted centroids, and improve the efficiency of memory usage. .

According to the inference apparatus 200 of the second embodiment, in the label prediction processing of CBCL, by reflecting the number of centroids in the correction coefficient when voting for the candidate class, the sample of the training data for each class It is possible to prevent the classification accuracy from deteriorating even when the number or feature distribution is biased. Especially when extending CBCL to cluster both the base class and the novel class, there is often a large bias in the number of samples and feature distributions between the base class and the novel class. Fairness can be ensured by correcting class votes.

The present invention has been described above based on the embodiments. It should be understood by those skilled in the art that the embodiments are examples, and that various modifications can be made to combinations of each component and each treatment process, and that such modifications are within the scope of the present invention. .

10 feature extraction unit 20 trained model storage unit 30 centroid generation unit 40 clustering unit 50 centroid storage unit 60 label prediction unit 70 neighborhood centroid selection unit 80 weighted voting unit 100 machine learning device , 200 reasoning apparatus.

Claims (5)

a feature extraction unit that extracts a feature vector of an input sample using a trained model;
a clustering unit that clusters the input samples to generate a centroid for each class based on the feature vector of the input samples and the class to which the input samples belong;
The clustering unit includes an initial session for clustering the base class training data set samples used to generate the trained model, and an additional session for clustering the newly added new class training data set samples, A machine learning device characterized by varying parameters used in clustering. The clustering unit selects a nearest neighbor centroid closest to the feature vector of the input sample from among a plurality of centroids of the class to which the input sample belongs, and the distance between the input sample and the nearest neighbor centroid is the parameter If the distance is less than a threshold, integrate the feature vector of the input sample into the nearest neighbor centroid, and if the distance is greater than or equal to the threshold, integrate the feature vector of the input sample into the centroid of the class to which the input sample belongs 2. The machine learning device according to claim 1, wherein the machine learning device is newly added as . 3. The machine learning device according to claim 2, wherein the clustering unit sets the threshold used in the initial session to be greater than the threshold used in the additional session. A feature extraction step of extracting a feature vector of the input sample using the trained model;
a clustering step of clustering the input samples to generate centroids for each class based on the feature vector of the input samples and the class to which the input samples belong;
The clustering step includes an initial session for clustering the base class training data set samples used in generating the trained model and an additional session for clustering the newly added new class training data set samples, A machine learning method characterized by varying parameters used in clustering. A feature extraction step of extracting a feature vector of the input sample using the trained model;
causing a computer to perform a clustering step of clustering the input samples to generate a centroid for each class based on the feature vectors of the input samples and the classes to which the input samples belong;
The clustering step includes an initial session for clustering the base class training data set samples used in generating the trained model and an additional session for clustering the newly added new class training data set samples, A machine learning program characterized by varying parameters used in clustering.

Images