JP7618917B2 - Machine learning device, machine learning method, and machine learning program - Google Patents

Description

The present invention relates to machine learning technology.

Humans are able to learn new knowledge through long-term experience, and are able to retain old knowledge without forgetting it. On the other hand, the knowledge of a Convolutional Neural Network (CNN) depends on the dataset used for training, and in order to adapt to changes in the data distribution, it is necessary to retrain the CNN parameters for the entire dataset. As a CNN learns new tasks, its estimation accuracy for old tasks decreases. Thus, when a CNN performs continuous training, it is unavoidable to suffer from catastrophic forgetting, in which the learning results of old tasks are forgotten while learning a new task.

Incremental learning or continual learning has been proposed as a method to avoid fatal forgetting. One method of continuous learning is PackNet.

Patent Document 1 discloses an information processing system that creates training data by labeling images that are similar to an input verification image, and trains a supervised image classifier using the created training data.

JP 2017-111731 A

PackNet, one method of continuous learning, can avoid the fatal forgetting problem. However, there is a problem in that if the input data does not have label information, it cannot be used for continuous learning. In addition, if the frequency with which the input data is used for learning is not controlled, bias in the learning may occur.

The present invention was made in light of these circumstances, and its purpose is to provide a machine learning technique that can assign label information to input data and control the frequency with which the input data is used for learning.

To solve the above problem, a machine learning device according to one embodiment of the present invention includes a classification unit that calculates the similarity between a feature map of an input task and a representative feature map of each task registered in advance, and assigns the label of the most similar task to the input task, a control unit that decides whether to use the input task for continued learning, and a continuous learning unit that, if the input task is to be used for continued learning, performs continuous learning using the labeled input task as training data.

Another aspect of the present invention is a machine learning method. This method includes a classification step of calculating the similarity between a feature map of an input task and a representative feature map of each task registered in advance, and assigning the label of the task with the highest similarity to the input task, a control step of determining whether or not to use the input task for continued learning, and a continuous learning step of performing continuous learning using the labeled input task as training data if the input task is to be used for continued learning.

In addition, any combination of the above components, and any transformation of the present invention into a method, device, system, recording medium, computer program, etc., are also valid aspects of the present invention.

The present invention provides a machine learning technique that can assign label information to input data and control the frequency with which the input data is used for learning.

FIG. 1 is a configuration diagram of a machine learning device according to an embodiment. FIG. 2 is a configuration diagram of a screening processing unit in FIG. 1. FIG. 3 is a configuration diagram of a screening unit in FIG. 2. FIG. 3 is a configuration diagram of a learning control unit in FIG. 2 . FIG. 2 is a configuration diagram of a task storage unit in FIG. 1 . FIG. 2 is a diagram showing the configuration of a task database in FIG. 1; 2 is a flowchart illustrating a continuous learning procedure by the machine learning device of FIG. 1 . FIG. 1 is a diagram illustrating a feature map output from each convolutional layer of a convolutional neural network. FIG. 1 is a diagram illustrating the concatenation of feature maps output from each convolutional layer of a convolutional neural network. FIG. 11 is a diagram illustrating a method for generating a concatenated feature map by adding feature maps with different numbers of channels. FIG. 13 is a diagram illustrating a method for generating a concatenated feature map by adding feature maps that differ in both the number of channels and the size. FIG. 1 illustrates a method for combining feature maps to generate a concatenated feature map.

Figure 1 is a configuration diagram of a machine learning device 100 according to an embodiment. The machine learning device 100 includes an input unit 10, a screening processing unit 20, a task storage unit 30, a continuous learning unit 40, a task database 50, an inference unit 60, and an output unit 70.

The input unit 10 supplies the tasks to be labeled to the screening processing unit 20 and supplies the unknown tasks to the inference unit 60. Here, as an example, the tasks are image recognition. For example, task 1 is the recognition of a specific object in an image, such as recognizing a cat, and task 2 is the recognition of a dog.

The screening processing unit 20 compares the feature map of the input task with the representative feature maps of tasks registered in the task database 50 to calculate the similarity, classifies the input task into the task with the highest similarity, and determines whether to use the input task for continued learning based on at least one of the frequency with which tasks similar to the input task were used for learning and the similarity.

When the input task image is input into a neural network model, which is an object recognition model, a feature map is output. If the feature map of the input image is similar to, for example, the feature map of a cat, the input image is recognized as an image of a cat and is labeled as a cat.

Here, the task does not recognize an object using only one feature map, but multiple feature maps. Therefore, the similarity is calculated by comparing a representative value, such as the average value, of the task's multiple feature maps with the feature map of the input task.

The similarity is calculated by comparing the absolute values of the elements of the feature maps. For example, if the size of the feature maps is 256 x 256, the absolute values of 65,536 elements are compared. A threshold is set here. If the similarity exceeds the threshold, the two feature maps are determined to overlap.

If each element of feature map A is _aij and each element of feature map B is _bij , the difference in absolute values between values at the same position between two feature maps A and B is calculated, for example, as _d1 (A,B), _d2 (A,B), _d∞ (A,B), and _dm (A,B) in the following equations.

In the above description, the similarity between the feature maps is calculated by calculating the difference between the absolute values of values at the same position between the two feature maps, but the similarity may be calculated by other methods. For example, for each feature map, the feature map absolute difference sum SAD is calculated as the sum of the horizontal absolute difference sum SAD_H and the vertical absolute difference sum SAD_V, and SAD = SAD_H + SAD_V is calculated. If the difference between the feature map absolute difference sum SAD_A of the feature map A and the feature map absolute difference sum SAD_B of the feature map B is smaller than a threshold value, the feature maps A and B may be determined to be overlapping. Here, for the sake of simplicity, assuming that the size of the feature map is 3 x 3, the elements of the first row of the feature map are a1, a2, a3, the elements of the second row are a4, a5, a6, and the elements of the third row are a7, a8, a9, the horizontal absolute difference sum SAD_H and the vertical absolute difference sum SAD_V are given by the following formulas.
SAD_H=|a1-a2|+|a2-a3|+|a4-a5|+|a5-a6|+|a7-a8|+|a8-a9|
SAD_V=|a1-a4|+|a2-a5|+|a3-a6|+|a4-a7|+|a5-a8|+|a6-a9|
As another method for calculating the similarity, a comparison of Euclidean distance or cosine distance may be used.

The task storage unit 30 updates the task database 50 with the classified input tasks, and controls whether to provide the classified input tasks to the continuous learning unit 40 based on at least one of the frequency and similarity. For example, if the frequency of the classified input task is less than the average frequency of the tasks registered in the task database 50, or if the similarity of the classified input task is less than a predetermined threshold, the classified input task is used for continuous learning, but if the frequency of the classified input task is equal to or greater than the average frequency, or if the similarity of the classified input task is equal to or greater than a predetermined threshold, the classified input task is not used for continuous learning. Here, the frequency of the input task is evaluated, for example, by the number of times the input task has been learned, or a relative index obtained by comparing the number of times the input task has been used with the number of times other tasks registered in the task database 50 have been used, or the like.

When the continuous learning unit 40 receives a classified input task from the task storage unit 30, it continues learning the object recognition model using the classified input task as supervised data.

The inference unit 60 uses the object recognition model learned by the continuous learning unit 40 to infer the classification of the unknown task received from the input unit 10. The output unit 70 outputs the inference result by the inference unit 60.

Figure 2 is a configuration diagram of the screening processing unit 20 in Figure 1. The screening processing unit 20 includes a screening unit 80 and a learning control unit 90.

The screening unit 80 compares the representative feature map of each task read from the task database 50 with the feature map of the input task, classifies the input task into the most similar task, and assigns label information to the input task. The screening unit 80 supplies the label information and similarity of the input task to the learning control unit 90, and supplies the feature map of the input task to the task storage unit 30.

The learning control unit 90 determines whether to provide the input task to the continuous learning unit 40 based on the frequency with which tasks with the same similarity as the input task have been learned.

Figure 3 is a configuration diagram of the screening unit 80. The screening unit 80 includes a representative feature map acquisition unit 82, a feature map conversion unit 84, and a classification unit 86.

The representative feature map acquisition unit 82 reads out the representative feature map of each task from the task database 50 and supplies it to the classification unit 86. The feature map conversion unit 84 calculates the feature map of the input task and supplies it to the classification unit 86. The classification unit 86 calculates the similarity between the feature map of the input task and the representative feature map of each task in the task database 50, classifies the input task into the task with the highest similarity, and assigns a label to the input task. The classification unit 86 supplies the label and similarity of the input task to the learning control unit 90, and supplies the feature map of the input task to the task storage unit 30.

Figure 4 is a configuration diagram of the learning control unit 90. The learning control unit 90 includes a frequency information acquisition unit 92, a similarity information acquisition unit 94, and a learning frequency control unit 96.

The frequency information acquisition unit 92 acquires the frequency at which a task with the same label and the same similarity as the input task is learned from the task database 50, and supplies this to the learning frequency control unit 96. The similarity information acquisition unit 94 supplies the similarity of the input task to the learning frequency control unit 96. In addition, the frequency information acquisition unit 92 supplies the frequency to the task storage unit 30, and the similarity information acquisition unit 94 supplies the similarity to the task storage unit 30.

The learning frequency control unit 96 determines whether or not to use the input task for continued learning based on at least one of the frequency and similarity, and notifies the task storage unit 30 of whether or not continued learning is to be used.

In continuous learning, the learning results are highly dependent on the order of the learning data, and any bias in the way the learning data is presented will have a significant impact on the results. To maintain the diversity of the learning data, we measure the frequency with which tasks with the same similarity are learned, and make the frequency distribution of the learning data uniform.

The learning frequency control unit 96 uses the input task for continued learning if the learning frequency of the input task is lower than the average learning frequency of each task registered in the task database 50, or if the similarity of the input task is lower than a predetermined threshold. This is because the inference accuracy of the entire task can be improved without bias by actively using tasks with low learning frequency or low similarity for continued learning.

For example, if cat images have been learned frequently and cat recognition is already sufficiently accurate, relearning the input cat images will result in biased learning. Therefore, by preventing additional learning after sufficient learning, the number of times learning with other tasks is made equal, lowering the risk of fatal forgetting. Conversely, if the frequency with which dog images are learned is low compared to other tasks, the number of times learning with other tasks is made equal by continuing to learn the input dog images.

Figure 5 is a configuration diagram of the task storage unit 30. The task storage unit 30 includes an update determination unit 32, which performs update determination for tasks registered in the task database 50.

The update determination unit 32 updates the tasks registered in the task database 50 as necessary based on the feature map, similarity, frequency, and whether or not continued learning is performed. The learning frequency of a task is stored for each similarity, and the frequency information is updated each time. The representative feature map of a task is updated using the feature map of the input task provided for continued learning. Input tasks that are not provided for continued learning are ignored and are not used to update the representative feature map.

The task storage unit 30 supplies the continued learning unit 40 with information on whether or not continued learning is to be performed. The continued learning unit 40 refers to the continued learning status, and if the input task is to be used for continued learning, it continues learning using the input task as teacher data, but if the input task is not to be used for continued learning, it ignores the input task.

Figure 6 is a diagram showing the configuration of the task database 50. The task database 50 stores a representative feature map 52 and a frequency 54 for each task. Here, the representative value of the feature map of a task is, for example, the average value, median value, mode value, maximum value, minimum value, etc. of multiple feature maps of the task.

Figure 7 is a flowchart explaining the continuous learning procedure by the machine learning device 100 of Figure 1.

The input unit 10 inputs a task with an unknown label and provides it to the screening processing unit 20 (S10).

In the screening processing unit 20, the classification unit 86 of the screening unit 80 classifies the input tasks based on the similarity between the feature map of the input task and the representative feature maps of each task registered in the task database 50, and assigns a label (S20).

In the screening processing unit 20, the learning frequency control unit 96 of the learning control unit 90 determines whether to use the input task for continued learning based on at least one of the frequency and similarity of the input task (S30).

The update determination unit 32 of the task storage unit 30 updates the frequency associated with the similarity in the task database 50, and if the input task is used for continuous learning, updates the representative feature map of the task registered in the task database 50 using the feature map of the input task (S40).

If the input task is to be used for continued learning (Y in S50), the continued learning unit 40 continues learning the object recognition model using the labeled input task (S60); if the input task is not to be used for continued learning (N in S50), the input task is ignored and the process proceeds to step S70.

If there are still tasks (N in S70), return to step S10 and input the next task. If the task is complete (Y in S70), end the continuous learning.

The feature map described in the above embodiment will be explained in detail with reference to Figures 8 to 12.

Figure 8 is a diagram explaining the feature maps output from each convolutional layer of a convolutional neural network. In this embodiment, the feature map refers to the output data of the convolutional layer when input data is input to the convolutional neural network. As shown in Figure 8, there are two types of feature maps: intermediate output data from the convolutional layers of each layer, and final output data from the convolutional layer of the final layer.

Figure 9 is a diagram explaining the concatenation of feature maps output from each convolutional layer of a convolutional neural network. When obtaining a feature map of input data, some features may be overlooked if only the feature map output from the convolutional layer of the final layer is used. Therefore, as shown in Figure 9, it is more preferable to create a feature map of the input data by concatenating the feature maps output from the convolutional layers of the intermediate layers as required and the final feature map output from the convolutional layer of the final layer. For example, concatenate feature map 2 of layer 2 and the final feature map of the final layer. Concatenate feature map 2 of layer 2, feature map 5 of layer 5, and the final feature map of the final layer. As shown above, there are various combinations, which can be used depending on the purpose.

Figure 10 is a diagram explaining a method for generating a concatenated feature map by adding feature maps with different numbers of channels. Here, a case will be explained in which feature map 2 with 64 channels and size 256 x 256 is concatenated with feature map 4 with 128 channels and size 256 x 256. To unify the number of channels, that is, to make the number of channels of feature map 2 the same as that of feature map 4, a convolutional layer is placed after feature map 2. This convolutional layer has 64 input channels, 128 output channels, and a filter size of 1 x 1. When feature map 2 is input to this convolutional layer, the size of new feature map 2 output from the convolutional layer does not change, but the number of channels is converted to 128. In other words, new feature map 2 has 128 channels and a size of 256 x 256. By adding new feature map 2 and feature map 4 with the unified number of channels, a concatenated feature map with 128 channels and a size of 256 x 256 is generated.

Figure 11 is a diagram explaining a method for generating a feature map by adding and concatenating feature maps with different numbers of channels and sizes. Here, a case will be explained in which a feature map 2 with 64 channels and a size of 256 x 256 is concatenated with a feature map 4 with 128 channels and a size of 64 x 64. To unify the number of channels, that is, to make the number of channels of the feature map 2 the same as that of the feature map 4, the number of channels of the feature map 2 is increased by zero padding. This is to fill the values of the added channels with zeros. As a result, the number of channels of the feature map 2 is converted to 128. After that, to unify the sizes, that is, to make the size of the feature map 2 the same as that of the feature map 4, the feature map 2 is downsampled. This is achieved by pooling with a stride of 4. Examples of pooling include maximum pooling, which outputs the maximum value, and average pooling, which outputs the average value, and are not limited to these, and other methods of pooling may be used. As a result, a new feature map 2 with 128 channels and a size of 64 x 64 is obtained. By adding new feature map 2 and feature map 4, which have the same number of channels and size, a concatenated feature map with 128 channels and a size of 64 x 64 is generated. Note that in order to make the number of channels in feature map 2 the same as that of feature map 4, a method of placing a convolutional layer after feature map 2 as described in Figure 10 may be adopted.

Figure 12 is a diagram explaining a method of generating a concatenated feature map by combining feature maps. Here, a case will be explained in which a feature map 2 with 64 channels and a size of 256x256 is concatenated with a feature map 4 with 128 channels and a size of 64x64. Since the feature maps have different sizes, the feature map 2 is downsampled to unify the sizes, i.e., to make the size of the feature map 2 the same as that of the feature map 4. This is achieved by pooling with a stride of 4. As a result, a new feature map 2 with 64 channels and a size of 64x64 is obtained. By concatenating the new feature map 2 and the feature map 4 with different numbers of channels and the same size, a concatenated feature map with 192 channels and a size of 64x64 is generated. Note that feature maps can be concatenated as a 3D tensor even if the number of channels is different, as long as they have the same size. For example, if feature map A has dimensions of [64 x 64 x 64] and feature map B has dimensions of [128 x 64 x 64] (where the size of the feature map is [H x W], the dimensions of the feature map are expressed as [number of channels x H x W]), when feature map A and feature map B are combined, if the sizes H x W are the same, the dimensions of the combined feature map will be [(number of channels A + number of channels B) x H x W] = [(64 + 128) x 64 x 64] = [192 x 64 x 64].

In FIG. 12, the feature maps are combined after unifying the sizes, but it is also possible to combine feature maps with the same number of channels but different sizes. For example, when a feature map A with 64 channels and a size of 64×64 is combined with a feature map B with 64 channels and a size of 128×64, a concatenated feature map with 64 channels and a size of 192×64 is generated. In general, when the size of the feature map is m×n and only one of m or n is the same between the two feature maps, the two feature maps can be combined. When the size of the feature map A is m× _n1 and the size of the feature map B is m× _n2 , the rows are aligned and each element of the feature map B is arranged to the left of each element of the feature map A, thereby generating a concatenated feature map with a size of m×(n ₁ +n ₂ ). If feature map A has size _m1 x n and feature map B has size _m2 x n, then a concatenated feature map of size ( _m1 + _m2 ) x n can be generated by aligning the columns and arranging each element of feature map B below each element of feature map A.

When calculating the similarity between the feature map of the input task concatenated in this way and the representative feature maps of each task registered in advance, the number of channels and size may differ. In this case, it is necessary to align both the number of channels and size using the two concatenation methods shown in Figures 10 and 11.

The various processes of the machine learning device 100 described above can of course be realized as a device using hardware such as a CPU and memory, but can also be realized by firmware stored in a ROM (read-only memory) or flash memory, or by software on a computer, etc. The firmware program or software program can be provided by recording it on a recording medium readable by a computer, etc., or it can be transmitted and received with a server via a wired or wireless network, or it can be transmitted and received as data broadcasting on terrestrial or satellite digital broadcasting.

As described above, according to the machine learning device 100 of this embodiment, it is possible to automatically label input tasks based on a feature map. In addition, by determining whether to use an input task for continued learning based on the learning frequency or similarity of the input task, it is possible to suppress learning bias and reduce the risk of fatal forgetting.

The present invention has been described above based on an embodiment. The embodiment is merely an example, and it will be understood by those skilled in the art that various modifications are possible in the combination of each component and each processing process, and that such modifications are also within the scope of the present invention.

10 Input unit, 20 Screening processing unit, 30 Task storage unit, 32 Update determination unit, 40 Continuous learning unit, 50 Task database, 52 Representative feature map, 54 Frequency, 60 Inference unit, 70 Output unit, 80 Screening unit, 82 Representative feature map acquisition unit, 84 Feature map conversion unit, 86 Classification unit, 90 Learning control unit, 92 Frequency information acquisition unit, 94 Similarity information acquisition unit, 96 Learning frequency control unit, 100 Machine learning device.

Claims (4)

a classification unit that calculates a similarity between a feature map of an input task and a representative feature map of each task registered in advance, and assigns a label of a task having the highest similarity to the input task;
a control unit that determines whether the input task is to be used for continuous learning;
a continuous learning unit that performs continuous learning using the labeled input task as training data when the input task is used for continuous learning,
The control unit determines whether to use the input task for further learning based on the frequency with which a task having a representative feature map that has the highest similarity to the feature map of the input task has been used for learning . The machine learning device according to claim 1 , wherein the control unit uses the input task for continuous learning when the frequency is less than a predetermined threshold. a classification step of calculating a similarity between a feature map of an input task and a representative feature map of each task registered in advance, and assigning a label of a task having the highest similarity to the input task;
a control step of determining whether the input task is to be used for further learning;
and a continuous learning step of performing continuous learning using the labeled input task as training data when the input task is used for continuous learning,
A machine learning method executed by a computer, characterized in that the control step determines whether to use the input task for further learning based on the frequency with which a task having a representative feature map that is most similar to the feature map of the input task has been used for learning . a classification step of calculating a similarity between a feature map of an input task and a representative feature map of each task registered in advance, and assigning a label of a task having the highest similarity to the input task;
a control step of determining whether the input task is to be used for further learning;
a continuous learning step of performing continuous learning using the labeled input task as training data when the input task is used for continuous learning ;
A machine learning program characterized in that the control step determines whether to use the input task for continued learning based on the frequency with which a task having a representative feature map that has the highest similarity to the feature map of the input task has been used for learning .

Images