Machine learning equipment, machine learning methods, and machine learning programs
The proposed machine learning device and method address catastrophic forgetting and improve classification accuracy by employing an attention attractor network and frequency domain analysis to regularize weights, enhancing feature extraction and classification performance in small-shot learning scenarios.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- JVC KENWOOD CORP
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-26
AI Technical Summary
Existing machine learning methods, such as Convolutional Neural Networks (CNNs), suffer from catastrophic forgetting, where learning new tasks leads to forgetting old tasks, and few-shot learning methods struggle with classification performance when data is scarce.
A machine learning device and method that utilizes an attention attractor network and frequency domain analysis to improve classification performance by regularizing weights of basic and new class classifiers, using local features and frequency domain information to enhance feature extraction and classification.
Improves classification accuracy by localizing input data features and utilizing frequency domain information, reducing catastrophic forgetting and enhancing recognition accuracy in small-shot learning scenarios.
Smart Images
Figure 2026086515000001_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to machine learning technology.
Background Art
[0002] Humans can learn new knowledge through long-term experience and can maintain the old knowledge without forgetting it. On the other hand, the knowledge of a Convolutional Neural Network (CNN) depends on the dataset used for learning, and in order to adapt to changes in the data distribution, it is necessary to relearn the parameters of the CNN for the entire dataset. In a CNN, as learning progresses for a new task, the estimation accuracy for the old task decreases. Thus, in a CNN, when continuous learning is performed, catastrophic forgetting, in which the learning results of the old task are forgotten during the learning of the new task, cannot be avoided. As a method for avoiding catastrophic forgetting, incremental learning or continual learning has been proposed. Incremental learning or continual learning is a learning method in which, when a new task or new data occurs, instead of learning the model from scratch, the currently learned model is improved and learned. On the other hand, since new tasks often have only a small number of sample data available, few-shot learning has been proposed as a method for efficiently learning with a small amount of teacher data. In few-shot learning, without relearning the parameters once learned, another
[0003]
[0004]
[0005] Learn a new task using a small number of parameters.
[0005] Learn a new class without fatal forgetting the learning results of the base class. Continuous learning and small-shot learning, which involves learning a small number of new classes compared to the basic class. This combination is called incremental few-shot learning (IFSL). A method has been proposed (Non-Patent Document 1). In continuous small-shot learning, the base class Therefore, we train from a large dataset, and for new classes, we use a small amount of sample data. You can learn from this. [Prior art documents] [Non-patent literature]
[0006] [Non-Patent Document 1] Ren, M., Liao, R., Fetaya, E., & Zemel, RS (2018). Incremental few-shot learning with attention attractor networks. arXiv preprint arXiv:1810.07218. [Non-Patent Document 2] Zhang, C., Cai, Y., Lin, G., & Shen, C. (2020). DeepEMD: Differentiable earth mover's distance for few-shot learning. arXiv preprint arXiv:2003.06777. [Non-Patent Document 3] Watanabe, T., & Wolf, DF (2021). Image classification in frequency domain with 2SReLU: a second harmonics superposition activation function. Applied Soft Computing, 112, 107851. [Overview of the project] [Problems that the invention aims to solve]
[0007] As a continuous small-shot learning method, the attention attractor network described in Non-Patent Document 1 ( There is an Attention Attractor Network. In an Attention Attractor Network, the basic class The performance of the classifier has a significant impact on new class classifiers and attention attractors. In the tractor network, when training the basic classifier, local features of images are used. Therefore, there was room for improvement in the classification performance of the basic classifier.
[0008] Furthermore, DeepEMD, described in Non-Patent Document 2, is a small-shot learning method. epEMD uses a distance measurement method called EMD (Earth Mover Distance) to measure the local area of an image. Image classification is performed by comparing local features. DeepEMD is used when the amount of training data is small. There was room for improvement in classification performance.
[0009] This invention has been made in view of these circumstances, and its purpose is to localize input data This provides machine learning techniques that can improve the classification performance of classifiers by utilizing their features. It is located there. Another purpose is to use the frequency domain information of the input data to improve the recognition accuracy of the learning process. The objective is to provide machine learning techniques that can improve [something].
Means for Solving the Problem
[0010] In order to solve the above problems, a machine learning device according to an aspect of the present invention is a machine learning device that continuously learns a small number of new classes compared to a basic class, and uses weights learned based on divided features of input data to extract features of the input data, a basic class classifier that classifies basic classes based on the features of the input data, a new class classifier that classifies new classes based on the features of the input data, and an attention attractor unit that regularizes the weights of basic class classification and the weights of new class classification.
[0011] Another aspect of the present invention is a machine learning method. This method is a machine learning method for continuously learning a small number of new classes compared to a basic class, and includes a feature extraction step of extracting features of the input data using weights learned based on divided features of the input data, a basic class classification step of classifying basic classes based on the features of the input data, a new class classification step of classifying new classes based on the features of the input data, and an attention attractor step of regularizing the weights of basic class classification and the weights of new class classification.
[0012] In addition, any combination of the above components, and those obtained by converting the expression of the present invention between a method, an apparatus, a system, a recording medium, a computer program, etc. are also effective as aspects of the present invention.
Effects of the Invention
[0013] According to an aspect of the present invention, by utilizing local features of input data, the classification performance of a classifier can be It can be improved. Furthermore, according to another aspect of the present invention, the frequency domain of the input data Information can be used to improve the accuracy of learning and recognition. [Brief explanation of the drawing]
[0014] [Figure 1] This diagram illustrates the configuration of the attention attractor network. [Figure 2] Figures 2(a) and 2(b) illustrate a two-stage training process. [Figure 3] This is a diagram showing the configuration of a machine learning system. [Figure 4] Figures 4(a) to 4(c) are diagrams that further break down the first stage of training shown in Figure 2(a) into three stages (stage 1.1, stage 1.2, and stage 1.3). [Figure 5] Figures 5(a) and 5(b) further break down the second stage of training shown in Figure 2(b) into two stages (stage 2.1 and stage 2.2). [Figure 6] Figure 3 is a diagram showing the configuration of the data control unit. [Figure 7] This diagram illustrates the configuration and operation of the learning unit in Figure 3 of Stage 1.1. [Figure 8] This diagram illustrates the configuration and operation of the learning unit in Figure 3 of the first and second stages. [Figure 9] This diagram illustrates the configuration and operation of the learning unit in Figure 3 of Stage 1.3. [Figure 10] This diagram illustrates the configuration and operation of the learning unit in Figure 3 of Stage 2.1. [Figure 11] This diagram illustrates the configuration and operation of the learning unit in Figure 3 of Stage 2.2. [Modes for carrying out the invention]
[0015] Figure 1 is a diagram illustrating the configuration of an attention attractor network. The twerk consists of a feature extraction unit 10, a basic class classification unit 20, a attention attractor unit 30, and a new class It includes a S classification unit 40 and a combined classification unit 50.
[0016] Attention attractor networks utilize Dynamic Few-Shot Learning. DFSL is one of the representative methods, and it uses attention attractors to form a base This class classification unit 20 and the new class classification unit 40 can be combined.
[0017] The feature extraction unit 10 includes a backbone CNN. The backbone CNN is a convolutional CNN. It is a CNN that extracts features from input data. The characteristics of the output input data are determined by the basic class classification unit 20, the attention attractor unit 30, and new It is supplied to the Class Classification Unit 40.
[0018] The basic class classification unit 20 includes a basic class classifier and basic class classification weights. Using the dataset as training data, the feature extraction unit 10 and the basic classification unit 20 are set together. Then pre-train. The base class dataset contains a large number of samples. An example of a weight is an image, but it is not limited to this. The basic classification weight is the basic classification This is the weight vector of the container, representing the average feature of the samples in the basic class dataset. That is the case.
[0019] The new class classification unit 40 includes a new class classifier and a new class for continuous learning of new classes. It includes the class classification weights. The new class classification unit 40 is post-trained on the new class dataset. The number of samples in the new class dataset is the same as in the base class dataset. It is a small number compared to the number of samples included. The new classification weights are the new classifier's This is a weight vector, which represents the average feature of the samples in the new class of dataset. ru.
[0020] The attention attractor unit 30 indicates that the basic class training data cannot be used in continuous learning. Under the given premise, when continuously training on a new class dataset, the base class dataset The learning process is controlled to prevent the forgetting of knowledge learned in the past and to avoid bias in inference results towards new classes.
[0021] The attention attractor unit 30 uses a common feature space to combine the basic classification weights and the new classification weights. Regularization is performed so that it can be compared above. Attention attractor unit 30 is from basic class classification unit 20 The provided base class information is used and added to the cross-entropy loss of the classification. The regularization term information is calculated and supplied to the new classification unit 40. The new classification unit 40 then Minimize the regularized cross-entropy loss, which is the sum of the cross-entropy loss and the regularization term. The new classification weights are updated accordingly.
[0022] The combined classification unit 50 combines the classification results from the basic class classification unit 20 with the new class classification unit 40. The classification results are combined and the final classification result is output.
[0023] In this embodiment, the feature extraction unit 10, the basic class classification unit 20, and the new class classification unit The discrimination performance is improved by improving 40 training methods, but attention attractors The training for the teamwork will be the same as the conventional method.
[0024] In this embodiment, the feature extraction unit 10 and basic class classification are performed based on the input data of the basic class. The first stage of training involves training unit 20, and the input data for the new class is used. The second stage of training involves training the new class classification unit 40 and the attention attractor unit 30. Perform the action.
[0025] Figures 2(a) and 2(b) illustrate the two-stage training. ) is the first stage of training, and the feature extraction unit 10 uses the input data of the basic class. The basic class classification unit 20 is then trained. Figure 2(b) shows the second stage of training. Yes, the weight of the feature extraction unit 10 and the basic class classification unit 20 learned in the first stage of training The unit is fixed, and the attention attractor unit 30 controls the basic class classification unit 20 and the new class classification unit 4 By associating 0s, the new class classification unit 40 and attention attractor use the input data of the new class. Train section 30.
[0026] Figure 3 is a diagram of the machine learning device 100. The machine learning device 100 includes an input unit 110, Data control unit 120, learning unit 130, weight storage unit 140, classification unit 150, and output unit 1 Includes 60.
[0027] The input unit 110 supplies the input training data to the data control unit 120. Unit 120 sorts the training data into base classes and new classes based on labels and then performs the learning process. Give it to 130.
[0028] Here, the configuration and operation of the data control unit 120 will be explained in detail with reference to Figure 6.
[0029] Figure 6 is a diagram showing the configuration of the data control unit 120. The data control unit 120 is a class classification unit 1 21. Basic class storage unit 122, new class storage unit 124, first basic class output unit 126 It includes a second basic class output unit 127 and a new class output unit 128.
[0030] The class classification unit 121 divides the input data into base classes and new classes based on labels. The data of the base class is stored in the base class storage unit 122, and the data of the new class is stored in the new class storage unit 122. Save to storage unit 124.
[0031] The first basic class output unit 126 outputs the basic class stored in the basic class storage unit 122. This process generates and outputs basic class training data for learning basic classes from force data.
[0032] The second basic class output unit 127 outputs the basic class stored in the basic class storage unit 122. Using force data, generate episodes consisting of a support set S and a queryset Q. And output.
[0033] The new class output unit 128 outputs the input data of the new class stored in the new class storage unit 124. Using the data, an episode consisting of a support set S and a query set Q is generated. Output.
[0034] Returning to Figure 3, let's describe the configuration of the machine learning device 100.
[0035] The learning unit 130 consists of the feature extraction unit 10 in Figure 1, the basic class classification unit 20, and the attention attractor unit 3. 0, and a new class classification unit 40, which are controlled to perform learning, and feature extraction unit 10 The weights of the basic class classification unit 20 and the new class classification unit 40 are stored in the weight storage unit 140. do.
[0036] The classification unit 150 consists of the feature extraction unit 10, the basic class classification unit 20, and the new class division shown in Figure 1. The class section 40 is controlled, and the basic class classification section 20 uses the weights stored in the weight storage section 140. The classification results of the new class classification unit 40 are combined and class classification is performed. The output unit 160 outputs the class Output the results of the classification.
[0037] Figures 4(a) to 4(c) show the first stage of training in Figure 2(a) further broken down into three stages. This diagram explains the process in stages (Stage 1.1, Stage 1.2, Stage 1.3).
[0038] Figure 4(a) shows the first stage of training. The feature extraction unit 10 extracts the basic class The training data is used as input data to extract features, which are then provided to the provisional basic class classification unit 22. The basic class classification unit 22 classifies the input data based on its features and outputs the classification result.
[0039] The provisional basic class classification unit 22 is a multilayer neural network which is a basic class classifier and Includes basic class classification weights. The basic class dataset is used as training data for the feature extraction unit. Pre-train the 10 and the provisional basic class classification unit 22 as a set.
[0040] Figure 7 is a diagram illustrating the configuration and operation of the learning unit 130 in the first stage. The output unit 126 inputs the basic class training data to the feature extraction unit 10. Feature extraction unit 1 0 extracts features from the input data and inputs them into the provisional basic class classification unit 22. The LASS classification unit 22 classifies the input data based on the input features. Classification result and true value By comparison, the weights of the feature extraction unit 10 and the provisional basic class classification unit 22 are adjusted to minimize the loss. Learn.
[0041] After the first stage of training, the provisional basic class classification unit 22 is discarded, but the first The weights of the feature extraction unit 10 learned in the first training stage are stored in the weight storage unit 140. Then, it is continued to the first and second stages.
[0042] Figure 4(b) shows the first and second stages of training. The feature extraction unit 10 extracts the basic class The support set S and query set Q are used as input data to extract features.
[0043] The division unit 90 divides the features output by the feature extraction unit 10 into multiple regions. Generally, images Information can be represented in the spatial domain or the frequency domain. Therefore, image features can be represented in the spatial domain or Divide it into the frequency domain.
[0044] When dividing into the frequency domain, for example, a cosine transformer can be used to extract features from the feature space into the frequency domain. The data is converted to a format, and the features are divided into multiple parts in the frequency domain. For example, from low-frequency components to high-frequency components. It is divided into six frequency domains up to . In the frequency domain, the data is obtained by combining frequency components. It is described as follows. High-frequency components are directly related to details and noise. In images, they define contours and boundaries. High-frequency components are important for interpretation. However, in some tasks such as image classification, The object's contour is not important information (Non-Patent Literature 3). Therefore, from low-frequency components to high-frequency components... By dividing the features into multiple frequency domains up to the wave component and using them for classification, the discrimination performance can be improved. It can be improved.
[0045] The distance calculation unit 80 calculates the similarity between the divided image features as a distance. EMD is, This is one method for calculating the similarity of image features as a distance, where two sets of feature quantities A and B are... Given the distance between the features of set A and the features of set B, feature sets A and B The distance between them can be calculated.
[0046] For example, using support set S and query set Q as input data, each can be used as an example. Then divide it into 9 regions.
[0047] The number of regions and the range of regions when dividing the frequency should be the same for the support set and the query set. In other words, if you divide the support set into three areas, the query set will also be divided into three areas. It needs to be divided. If we divide it into three areas, the scope of each area would be, for example, support sets and quests. With a reset, the range is 0-10Hz, 10-20Hz, and 10-30Hz. The frequency range to be considered is 0 to 30 Hz, and frequencies above 30 Hz will be excluded from the calculation. While preferable, it is also acceptable to divide the range above 30Hz and use it for distance calculations.
[0048] The number of frequency division regions and the range of the regions may always be the same for each task, but they may also be different. Good. The above shows examples of domain scopes in the case of three domains, but for other tasks, for example... Alternatively, the number of regions could be set to 2, and the divisions could be 0-15Hz and 15-30Hz. If the number of regions and the range of regions are the same for the port set and query set, then for each task the number of regions and The scope of the area may be determined on a case-by-case basis.
[0049] Distance calculations are performed between each region of support set S and each region of query set Q, and the calculation results are as follows. The results are combined and used as the loss. The classifier parameters are learned to minimize this loss. Performing distance calculations between each region of port set S and each region of query set Q means, for example, If we denote the areas of port set S as A and B, and the areas of query set Q as C and D, then A and C, and A and This involves calculating the distances between D, B and C, and B and D, and then combining the results of these distance calculations. That is the loss. Specifically, the DeepEMD rangefinder described in Non-Patent Literature 2 Measuring instruments can be used.
[0050] Based on the distance calculation result output by the distance calculation unit 80, the input is classified and the classification result is The true value is compared with the given value, and the weights of the feature extraction unit 10 are learned to minimize the loss.
[0051] Figure 8 is a diagram illustrating the configuration and operation of the learning unit 130 in the first and second stage. Stage 1.1 The weights of the feature extraction unit 10 that have been trained are read from the weight storage unit 140, and the first and second stages of feature extraction are performed. This value is reused as the initial value for the weight of the characteristic extraction unit 10.
[0052] The second base class output unit 127 outputs from the base class support set S and query set Q. Episode-formatted input data is input to the feature extraction unit 10. The feature extraction unit 10 processes the input Features are extracted from the data, and the features, divided into multiple regions, are input to the distance calculation unit 80. The calculation unit 80 calculates the data of the support set divided into multiple frequency domains and the multiple frequency domains Distance calculations are performed between the data in the queryset divided into regions, and the results of the distance calculations are combined. The loss is then calculated. The weights of the feature extraction unit 10 are retrained to minimize the loss. Section 1.2 The weights of the feature extraction unit 10, learned in stages, are stored in the weight storage unit 140.
[0053] Figure 4(c) shows the first and third stage of training. The feature extraction unit 10 extracts the basic class The training data is used as input data to extract features. The basic classification unit 20 uses random weighting Under initial values, representative features that represent the characteristics of all classes of the base class without input data. Outputs.
[0054] The division unit 90 divides the features extracted by the feature extraction unit 10 into multiple regions and performs basic class classification. The representative features output by section 20 are divided into multiple regions.
[0055] The distance calculation unit 80 uses the multiple features extracted by the feature extraction unit 10 and the basic class classification unit 20 The distance between multiple representative features output is calculated and combined, and the class with the smallest combined distance is selected. The input data is classified into categories. The classification result is compared with the true value, and the basic logic is used to minimize loss. The weights of the las classification unit 20 are learned.
[0056] Figure 9 is a diagram illustrating the configuration and operation of the learning unit 130 in the 1.3 stage. 1.2 stage The weights of the feature extraction unit 10 that have been trained are read from the weight storage unit 140, and the 1.3 stage features This is set as the weight for the feature extraction unit 10. In the 1.3 stage, the weight for the feature extraction unit 10 is learned. It is fixed. Random initial values are set for the weights of the basic class classification unit 20, and the basic class The features output from the Wrath classification unit 20 will be made to resemble the features extracted from the feature extraction unit 10. The weights of the basic class classification unit 20 are then learned.
[0057] The first basic class output unit 126 inputs the basic class training data to the feature extraction unit 10. The feature extraction unit 10 extracts features from the input data of the basic class and divides the features into multiple regions. The data is divided and given to the distance calculation unit 80. The basic class classification unit 20 calculates the characteristics of all classes of the basic class. The system outputs representative features, divides these representative features into multiple regions, and provides them to the distance calculation unit 80. The distance calculation unit 80 uses the multiple features extracted by the feature extraction unit 10 and the basic class classification unit 20. The distance between the multiple representative features output by is calculated and combined, and this is provided to the classification unit 150. 150 classifies the input data based on the combined distance calculation results and compares the classification result with the true value. By comparison, the weights of the basic class classification unit 20 are learned to minimize the loss.
[0058] Figures 5(a) and 5(b) show the second stage of training in Figure 2(b) in more detail. This diagram explains the process in two stages (Stage 2.1 and Stage 2.2). Stage 2.1 is about The second stage of internal learning is called external learning.
[0059] Figure 5(a) shows the training for Stage 2.1. In Stage 2.1, attention attraction The weight of section 30 is fixed, and the new class classification section 40 is created using the support set S of the new class. Train the weights.
[0060] The feature extraction unit 10 extracts features using the support set S of the new class as input data. The new class classification unit 40, under the initial value of random weights, performs a new class classification without input data. Outputs representative characteristics that represent all classes of Lass.
[0061] The division unit 90 divides the features extracted by the feature extraction unit 10 into multiple regions and performs new class classification. The representative features output by section 40 are divided into multiple regions.
[0062] The distance calculation unit 80 uses the multiple features extracted by the feature extraction unit 10 and the new class classification unit 40 The distance between multiple representative features output is calculated and combined, and the class with the smallest combined distance is selected. The input data is classified into categories. Pay attention to the difference between the classification result and the true value. Add the output value of the attractor unit 30. The loss is calculated, and the weights of the new classification unit 40 are learned to minimize the loss.
[0063] Figure 10 is a diagram illustrating the configuration and operation of the learning unit 130 in stage 2.1. Feature extraction unit 10 is set to the trained weights stored in the weight storage unit 140 in the first stage. The class classification unit 20 has also been trained in the first stage. The weights of the new class classification unit 40 include A random initial value is set.
[0064] In internal learning, the new class output unit 128 features a support set S for the new class. The data is input to the extraction unit 10. The feature extraction unit 10 extracts features from the input data of the new class. The features are divided into multiple regions and provided to the distance calculation unit 80. The new class classification unit 40 then calculates the new class Output representative features that represent the characteristics of all classes of Lass, and divide these representative features into multiple regions. This is provided to the distance calculation unit 80. The distance calculation unit 80 uses the multiple features extracted by the feature extraction unit 10 and The new classification unit 40 calculates the distance between multiple representative features output by the new class classification unit 40 and combines them. The attractor unit 30 pays attention to the output value of the distance calculation unit 80 and adds the result of the attractor network. The calculation determines the loss. The classification unit 150 outputs the classification result and implements a new method to minimize the loss. Learn the standard class classification unit 40.
[0065] Figure 5(b) shows the training in Stage 2.2. In Stage 2.2, new classes The weights of the class section 40 are fixed, and the attention attractor section 30 is calculated using the queryset Q of the new class. Train with weights.
[0066] The feature extraction unit 10 extracts features using the queryset Q of the new class as input data. The new class classification unit 40 is a representative that represents the characteristics of all classes of the new class without input data. Output the features.
[0067] The division unit 90 divides the features extracted by the feature extraction unit 10 into multiple regions and performs new class classification. The representative features output by section 40 are divided into multiple regions.
[0068] The distance calculation unit 80 uses the multiple features extracted by the feature extraction unit 10 and the new class classification unit 40 The distance between multiple representative features output is calculated and combined, and the class with the smallest combined distance is selected. The input data is classified into categories. Pay attention to the difference between the classification result and the true value. Add the output value of the attractor unit 30. The loss is calculated, and the weights of the attractor unit 30 are carefully learned to minimize the loss.
[0069] Figure 11 is a diagram illustrating the configuration and operation of the learning unit 130 in stage 2.2. Feature extraction unit In the first stage, the learned weights stored in the weight storage unit 140 are set to 10.
[0070] In outer learning, the new class output unit 128 extracts features from the queryset Q of the new class. Input to output unit 10. Feature extraction unit 10 extracts features from the input data of the new class. The characteristics are divided into multiple regions and provided to the distance calculation unit 80. The new class classification unit 40 then calculates the new class It outputs representative features that represent the characteristics of all classes, divides the representative features into multiple regions, and then outputs the distance This is provided to the distance calculation unit 80. The distance calculation unit 80 uses the multiple features extracted by the feature extraction unit 10 and new The distance between multiple representative features output by the classification unit 40 is calculated and combined. The tractor unit 30 pays attention to the output value of the distance calculation unit 80 and adds the result of the attractor network. Then, the loss is calculated. The classification unit 150 outputs the classification result and takes care to minimize the loss. Learn about the attractor unit 30.
[0071] Stages 2.1 and 2.2 are one-episode training sessions. Generally, the episodes are Since there are several hundred items, stages 2.1 and 2.2 are repeated several hundred times. Inner learning occurs each time. When this happens, the new class classification unit 40 is trained from initial values, but the attention attractor unit 30 The weights are carried over from the previous weights and trained accordingly.
[0072] The various processes of the machine learning device 100 described above utilize hardware such as the CPU and memory. Of course, it can be realized using ROM (Read-Only Memory). Firmware stored in flash memory, etc., and software on computers, etc. This can also be achieved through wearables. Specifically, through firmware programs and software. The program may also be provided by recording it on a recording medium that can be read by a computer, etc. Alternatively, data can be sent and received to and from the server via a wireless network, or via terrestrial or satellite radio. It is also possible to transmit and receive data broadcasts as digital broadcasting.
[0073] As described above, according to this embodiment, the characteristics of the input data are divided into multiple regions, By classifying input data using multiple local features, the accuracy of classification can be improved. This is possible. In addition, by dividing the features of the input data into multiple frequency domains, the recognition accuracy can be improved. It can improve.
[0074] The present invention has been described above based on embodiments. The embodiments are illustrative and their respective components The fact that various variations are possible in the combination of constituent elements and each processing process, and such variations Those skilled in the art will understand that this also falls within the scope of the present invention. [Explanation of Symbols]
[0075] 10 Feature extraction unit, 20 Basic class classification unit, 22 Provisional basic class classification unit, 3 0 Attention Attractor Unit, 40 New Classification Unit, 50 Combined Classification Unit, 80 Distance Calculation unit, 90 division unit, 100 machine learning device, 110 input unit, 120 data Control unit, 121 Class classification unit, 122 Basic class storage unit, 124 New class Storage unit, 126 First basic class output unit, 127 Second basic class output unit, 12 8 New class output unit, 130 Learning unit, 140 Weight storage unit, 150 Classification unit, 160 Output section.
Claims
1. A machine learning device that continuously learns a small number of new classes compared to the base class, Using weights learned based on the segmented features of the input data, the features of the input data are... A feature extraction unit that extracts characteristics, A division unit that divides the characteristics of the input data into multiple frequency domains, The data of the support set divided into multiple frequency domains by the division unit and the division The distance calculation is performed between the queryset data, which is divided into multiple frequency domains, by the section. , a distance calculation unit that synthesizes the results of distance calculations, The feature extraction unit minimizes the result of the distance calculation synthesized by the distance calculation unit. A machine learning device characterized by including a learning unit that retrains weights.
2. This is a machine learning method that involves continuous learning of a small number of new classes compared to the base class. Using weights learned based on the segmented features of the input data, the features of the input data are... A feature extraction step to extract characteristics, A division step that divides the characteristics of the input data into multiple frequency domains, The data of the support set divided into multiple frequency domains by the above division step and The distance between the data in the queryset, which has been divided into multiple frequency domains by the division step, is A distance calculation step that performs distance calculation and synthesizes the results of the distance calculation, To minimize the result of the distance calculation synthesized by the distance calculation step, the feature extraction A computer that includes a learning step in which it relearns the weights of the outgoing step is executed. A machine learning method.
3. A machine learning program that continuously learns a small number of new classes compared to the base class, Using weights learned based on the segmented features of the input data, the features of the input data are... A feature extraction step to extract characteristics, A division step that divides the characteristics of the input data into multiple frequency domains, The data of the support set divided into multiple frequency domains by the above division step and The distance between the data in the queryset, which has been divided into multiple frequency domains by the division step, is A distance calculation step that performs distance calculation and synthesizes the results of the distance calculation, To minimize the result of the distance calculation synthesized by the distance calculation step, the feature extraction This method is characterized by having the computer perform a learning step in which it relearns the weights of the previous steps. A machine learning program.