JP2020119048A - DNN selection program, DNN selection method, and information processing apparatus - Google Patents

Abstract

To select a more appropriate DNN for new data when there are a plurality of DNNs.SOLUTION: An information processing apparatus 1 learns, for each of a plurality of DNNs, a restorer that restores data input to the DNN from characteristics obtained from an intermediate layer of the DNN based on data different from the data used for learning each of the plurality of DNNs. The information processing apparatus 1 selects, for new data to be determined, a DNN corresponding to the restorer selected based on the data output from each of the plurality of restorers and the new data to be determined.SELECTED DRAWING: Figure 6B

Description

The present invention relates to a DNN selection program and the like.

A technique for discriminating (classifying) input data using a deep learning neural network (DNN) is known (for example, see Patent Document 1).

In a situation where a plurality of learned DNNs exist, it may be desired to determine which DNN is optimal for data to be newly discriminated.

In such a case, the information processing apparatus can determine an appropriate DNN for the new data based on the certainty factor of the label output from each of the plurality of DNNs.

International Publication No. 2016/132468

However, when there are a plurality of DNNs, it is difficult to select a more suitable DNN for new data.

For example, when the assumption that the confidence level of any label (class) with respect to unknown data is low holds, the information processing apparatus can select the DNN based on the confidence level of the label output by the DNN. However, it is known that the selection based on the certainty factor of the label may not be appropriate depending on the relationship between the data to be discriminated and the DNN. That is, since the DNN often outputs a high degree of certainty even for unknown data, it may not be possible to select a more appropriate DNN for new data.

The present invention, in one aspect, aims to select a more suitable DNN for new data when a plurality of DNNs exist.

In one aspect, the DNN selection program causes a computer to restore, for each of a plurality of DNNs (Deep Learning Neural Networks), a restorer that restores data input to the DNN from a feature obtained from an intermediate layer of the DNN. Learning is performed using data different from the data used for learning each DNN, and for specific data, a DNN corresponding to the data output from each of the plurality of restorers and the restorer selected based on the specific data is selected. Let the process run.

According to one embodiment, when there are a plurality of DNNs, it becomes possible to select a more suitable DNN for new data.

FIG. 1 is a functional block diagram of the configuration of the information processing apparatus according to the first embodiment. FIG. 2 is a diagram illustrating an idea of DNN selection according to the first embodiment. FIG. 3 is a diagram illustrating an example of the flow of the restorer learning according to the first embodiment. FIG. 4 is a diagram illustrating an example of the flow of DNN selection according to the first embodiment. FIG. 5A is a diagram illustrating an example of layer determination. FIG. 5B is a diagram showing another example of layer determination. FIG. 6A is a diagram (1) showing a comparison of the operation of the DNN selection between the first embodiment and the conventional technique. FIG. 6B is a diagram (2) showing a comparison of the operation of the DNN selection between the first embodiment and the conventional technique. FIG. 7 is a diagram illustrating an example of a flowchart of the DNN selection process according to the first embodiment. FIG. 8 is a functional block diagram of the configuration of the information processing apparatus according to the second embodiment. FIG. 9 is a diagram illustrating an example of the flow of the restorer learning and the auxiliary DNN learning according to the second embodiment. FIG. 10 is a diagram illustrating an example of the flow of DNN selection according to the second embodiment. FIG. 11 is a diagram illustrating an example of a flowchart of the DNN selection process according to the second embodiment. FIG. 12 is a diagram illustrating an example of a computer that executes a DNN selection program.

Hereinafter, embodiments of the DNN selection program, the DNN selection method, and the information processing device disclosed in the present application will be described in detail with reference to the drawings. The present invention is not limited to the examples.

[Configuration of Information Processing Device According to First Embodiment]
FIG. 1 is a functional block diagram of the configuration of the information processing apparatus according to the first embodiment. When there are a plurality of learned DNNs, the information processing apparatus 1 illustrated in FIG. 1 selects which DNN to apply to new data to be determined.

Here, the idea of the DNN selection performed by the information processing apparatus 1 according to the first embodiment will be described with reference to FIG. FIG. 2 is a diagram illustrating an idea of DNN selection according to the first embodiment. In FIG. 2, it is assumed that DNN-1 is the DNN that determines the type of dog that has been learned. The DNN-1 discriminates the type of dog from the discrimination target data, and outputs the certainty factor for each label (here, the dog type).

As shown in FIG. 2, DNN-1 generally learns to extract general features in a shallow layer and specific features required to distinguish a dog type in a deep layer. That is, in the shallow layer, general features (such as edges) are extracted from the image, and in the deep layer, features necessary for distinguishing the type of dog that the DNN is in charge of are extracted, and the dog type that the DNN is in charge of is extracted. Information that is not needed to determine is considered missing. For example, in the deep layer of DNN-1 that discriminates the type of dog, it is considered that almost no features necessary for discriminating the type of cat different from the dog remain.

Therefore, the information processing device 1 is appropriate for selecting the DNN-1 depending on how much information necessary for the discrimination in the deeper layer of the DNN remains for the new data to be discriminated. Estimate. That is, the information processing device 1 selects a more appropriate DNN for the discrimination with respect to the new data to be discriminated.

In order to select an appropriate DNN from a plurality of DNNs, the information processing device 1 has a restorer that restores the data input to the DNN from the characteristics obtained from the intermediate layer (deep layer) of the DNN for each of the plurality of DNNs. Try to learn. Then, the information processing apparatus 1 selects the data output from each of the plurality of restorers and the DNN corresponding to the restorer selected based on the new data, for the new data to be determined.

Returning to FIG. 1, the information processing device 1 has a control unit 10 and a storage unit 20.

The control unit 10 corresponds to an electronic circuit such as a CPU (Central Processing Unit). The control unit 10 has an internal memory for storing programs and control data that define various processing procedures, and executes various processing by these. The control unit 10 includes a restoration device learning unit 11, a restoration data generation unit 12, a DNN selection unit 13, a determination unit 14, and a layer determination unit 15. The restorer learning unit 11 is an example of the restorer learning unit. The DNN selection unit 13 is an example of the DNN selection unit.

The storage unit 20 is, for example, a RAM, a semiconductor memory device such as a flash memory, or a storage device such as a hard disk or an optical disk. The storage unit 20 has a data group 21, restored data 22, and restored data (new data) 23.

The data group 21 is a set of learning data such as a general natural image used for learning the restorer. The data group 21 is a set of learning data different from the learning data of each DNN.

The restored data 22 is the data output from the restorer when learning the restorer that restores the data input to the DNN from the characteristics obtained from the intermediate layer of the DNN, and is the data for each of the plurality of DNNs. ..

The restored data (new data) 23 is the data output from each of the plurality of restorers for the new data to be determined.

The restorer learning unit 11 learns the restorer corresponding to the DNN for each of the plurality of DNNs. The term "restoring device" as used herein refers to a learning device that restores the data input to the DNN from the characteristics obtained from the intermediate layer (deep layer) of the DNN. The restorer learns using general data, but learns using data different from the data used for learning each DNN. This is to ensure independence between the data used for learning the DNN and the data used for learning the restorer.

For example, the restorer learning unit 11 inputs the data included in the data group 21 to a plurality of DNNs. The restorer learning unit 11 extracts the respective features obtained from the intermediate layers of the plurality of DNNs and inputs the extracted respective features to the restorers corresponding to the plurality of DNNs. The restorer learning unit 11 compares the restored data 22 output from each restorer with the data input to each DNN, and learns each restorer so that the error becomes small. The intermediate layer from which the feature is extracted is determined by the layer determination unit 15 described later.

The restored data generation unit 12 generates restored data 22 for the new data to be determined through the DNN and the restorer. That is, the restored data generation unit 12 inputs new data to be determined to the DNN, inputs the features extracted from the middle layer of the DNN to the restorer, and outputs the restored data (new data) 23 from the restorer. ..

The DNN selection unit 13 selects the restorer with the smallest error based on the data output from each of the plurality of restorers and the new data for the new data, and selects the DNN corresponding to the selected restorer. If the error is the smallest, it means that the intermediate layer of the corresponding DNN has more effective features, so the DNN selecting unit 13 selects the DNN as a DNN for discriminating new data. That is, the DNN selection unit 13 estimates how many effective features each DNN has extracted from the new data due to the error. As a result, the confidence factor is often high for unknown data input to the DNN, while the error can be expected to be large for unknown data input to the DNN.

The discriminating unit 14 discriminates (classifies) new data to be discriminated using the selected DNN.

Here, an example of the flow of the restorer learning by the restorer learning unit 11 will be described with reference to FIG. FIG. 3 is a diagram illustrating an example of the flow of the restorer learning according to the first embodiment. FIG. 3 shows a plurality of DNNs and a restorer corresponding to each of the plurality of DNNs. As the plurality of DNNs, DNN-1 for discriminating the type of dog, DNN-2 for discriminating the type of cat, and DNN-3 for discriminating the type of bird are listed. The restorer-1 is mentioned as a restorer corresponding to DNN-1. The restorer-2 is mentioned as a restorer corresponding to DNN-2. The restorer-3 is mentioned as a restorer corresponding to DNN-3.

Under such a situation, the restorer learning unit 11 inputs the data included in the data group 21 to DNN-1, DDN-2, and DNN-3. The restorer learning unit 11 extracts the features obtained from the intermediate layer of DNN-1 and inputs the extracted features to the restorer-1. The restorer learning unit 11 extracts the features obtained from the intermediate layer of DNN-2 and inputs the extracted features to the restorer-2. The restorer learning unit 11 extracts a feature obtained from the intermediate layer of DNN-3 and inputs the extracted feature to the restorer-3. The restorer learning unit 11 compares the restored data 22-1 output from the restorer-1 with the data input to the DNN-1, and learns the restorer-1 so that the error becomes small. The restorer learning unit 11 compares the restored data 22-2 output from the restorer-2 with the data input to the DNN-2, and learns the restorer-2 so that the error becomes small. The restorer learning unit 11 compares the restored data 22-3 output from the restorer-3 with the data input to the DNN-3, and learns the restorer-3 so that the error becomes small.

Here, an example of the flow of DNN selection by the DNN selection unit 13 will be described with reference to FIG. FIG. 4 is a diagram illustrating an example of the flow of DNN selection according to the first embodiment. FIG. 4 shows a plurality of DNNs and a learned restorer corresponding to each of the plurality of DNNs. As the plurality of DNNs, DNN-1 for discriminating the type of dog, DNN-2 for discriminating the type of cat, and DNN-3 for discriminating the type of bird are listed. The restorer-1 is mentioned as a restorer corresponding to DNN-1. The restorer-2 is mentioned as a restorer corresponding to DNN-2. The restorer-3 is mentioned as a restorer corresponding to DNN-3.

Under such a situation, the restored data generation unit 12 inputs new data to be determined to DNN-1, DNN-2 and DNN-3. The restored data generation unit 12 extracts the features obtained from the intermediate layer of DNN-1, inputs the extracted features into the restorer-1, and outputs the restored data-1. The restored data generation unit 12 extracts the features obtained from the intermediate layer of DNN-2, inputs the extracted features to the restorer-2, and outputs the restored data-2. The restored data generation unit 12 extracts the features obtained from the intermediate layer of the DNN-3, inputs the extracted features to the restorer-3, and outputs the restored data-3. The DNN selection unit 13 compares the restored data-1 output from the restorer-1 with the new data, and calculates an error. The DNN selection unit 13 compares the restored data-2 output from the restorer-2 with the new data, and calculates an error. The DNN selection unit 13 compares the restored data-3 output from the restorer-3 with the new data, and calculates an error.

Then, the DNN selection unit 13 selects the restorer with the smallest error and selects the DNN corresponding to the selected restorer. Here, the error between the restored data-1 and the new data is calculated as 0.5. The error between the restored data-2 and the new data is calculated as 0.1. The error between the restored data-3 and the new data is calculated as 1.2. Therefore, the DNN selection unit 13 selects the restorer-2 that outputs the restored data-2 having the smallest error, and selects the DNN-2 corresponding to the selected restorer-2.

That is, the DNN selection unit 13 estimates whether or not the selection of each DNN is appropriate, depending on how much information necessary for the discrimination remains in the middle layer of the DNN for the new data to be discriminated. As a result, the DNN selection unit 13 can select a more appropriate DNN for the discrimination with respect to the new data to be discriminated.

Returning to FIG. 1, the layer determination unit 15 determines the intermediate layer from which the features of DNN are extracted.

In one example, when the DNN is a convolutional NN (Neural Network), the layer determination unit 15 determines the layer before the fully connected layer as the intermediate layer from which the features of the DNN are extracted. The front layer is, for example, the previous layer. The convolutional NN as used herein refers to a DNN having a configuration in which a convolutional layer and a pulling layer are repeated, and then a total coupling layer is continued.

Here, an example of layer determination by the layer determination unit 15 will be described with reference to FIG. 5A. FIG. 5A is a diagram illustrating an example of layer determination. It is assumed that DNN-1 is represented as the convolutional NN in FIG. 5A. For example, the layer determination unit 15 determines the layer immediately before the fully connected layer of DNN-1 as the intermediate layer from which the features of DNN are extracted.

Further, in another example, when the restorer is configured for each layer of the DNN, the layer determining unit 15 performs the following processing. That is, the layer determination unit 15 determines, for each of the restorers, the layer for the restorer that maximizes the variance of the restore error for the data group 21 used for learning the restorer, as the intermediate layer for extracting the DNN feature. ..

Here, another example of layer determination by the layer determination unit 15 will be described with reference to FIG. 5B. FIG. 5B is a diagram showing another example of layer determination. FIG. 5B shows a DNN-1 and a reconstructor configured for each layer of DNN-1. For example, the layer determination unit 15 determines, for each restorer, the layer for the restorer that maximizes the variance of the restore error for the data group 21 used for learning the restorer, as the intermediate layer for extracting the DNN feature. .. The reason why the layer for the restorer that maximizes the variance of the restoration error is determined as the intermediate layer is as follows. If the variance of the restoration error is large, there are data that can be restored well and data that cannot be restored well. Therefore, there is a difference in the quality of restoration between the class that is good at restoration and the class that is not good at restoration, and it is easy to use for the selection of DNN.

Here, a comparison of the operation of the DNN selection between the first embodiment and the conventional technique will be described with reference to FIGS. 6A and 6B. 6A and 6B are diagrams showing a comparison of the operation of the DNN selection between the first embodiment and the conventional technique. FIG. 6A is a diagram showing a conventional DNN selection operation. FIG. 6A shows DNN-1 for discriminating the type of dog and DNN-2 for discriminating the type of cat. Here, it is assumed that a cat image is input to each DNN as new data.

When the image of the cat is discriminated by DNN-1, the discrimination result is mapped to the class of dog breed c in the feature space of DNN-1, but from the boundary surface of each class of dog breed abc. It is mapped to a distant place and a place distant from the already learned learning data. This is because the cat image is unknown data for DNN-1, which discriminates the type of dog, and is therefore mapped to a place apart from the already learned learning data. However, since the determination result is far from the boundary surface, DNN-1 determines that the breed is c with a high degree of certainty.

When the image of the cat is discriminated by DNN-2, the discrimination result is mapped to the class of cat species a in the feature space of DNN-2, but on the boundary surface of each class of cat species abc. It is mapped to a place close to the same place as the already learned learning data. This is because the DNN-2 that discriminates the type of cat has seen a cat image that is new data, and is therefore mapped to the same place as the already learned learning data. However, since the determination result is close to the boundary surface, DNN-2 determines that it is a cat breed a with a low certainty factor.

Therefore, in the conventional DNN selection, since the new data is a cat image, DNN-2 should be selected originally, but DNN-1 that outputs a higher certainty factor is selected.

On the other hand, FIG. 6B is a diagram showing the operation of the DNN selection according to the first embodiment. 6B shows DNN-1 and DNN-2 similar to FIG. 6A. Furthermore, the restorer-1 is shown as a restorer corresponding to DNN-1. A restorer-2 is shown as a restorer corresponding to DNN-2. The restorers-1 and 2 have already been learned by the restorer learning unit 11 so that the error between the data input to the DNN-1 and DNN and the restored data becomes small. Here, it is assumed that a cat image is input to each DNN as new data.

The DNN selection unit 13 according to the first embodiment extracts the features obtained from the intermediate layer of the DNN-1, inputs the extracted features to the restorer-1, and outputs the restored image output from the restorer-1 and a new image. The error is calculated by comparing the data with the image of the cat which is various data. Here, since DNN-1 only learns to discriminate dog breeds, the characteristics of cats do not remain in the middle layer. Therefore, the DNN selection unit 13 calculates a large value as the error. That is, when the image of the cat is discriminated by DNN-1, the discrimination result is already learned in the feature space of DNN-1 at a place distant from the boundary surface of each class of dog breed abc. It is mapped to a place apart from the learning data. In the conventional technique, the certainty factor of the dog breed c is higher than the cat image which is the unknown image, but the DNN selecting unit 13 according to the first embodiment restores the error of the cat image which is the unknown image. Will grow.

The DNN selection unit 13 extracts the features obtained from the intermediate layer of the DNN-2, inputs the extracted features to the restorer-2, and outputs the restored image output from the restorer-2 and the cat that is new data. Then, the error is calculated by comparing with the image of. Here, since DNN-2 is learning to discriminate the cat species, the features regarding cats remain in the middle layer. Therefore, the DNN selection unit 13 calculates a smaller value as an error than in the case of the restorer-1. That is, when a cat image is discriminated by DNN-1, the discrimination result is mapped to the class of cat species a, but it is already learned at a place near the boundary surface of each class of cat species abc. It is mapped to the same place as the training data. In the conventional technique, the certainty factor of the cat species a is lower than that of the cat image that is mapped close to the boundary surface. However, the DNN selection unit 13 according to the first embodiment performs the learning of the cat image with respect to the cat image. The restoration error becomes small.

Therefore, the DNN selection unit 13 according to the first embodiment can select the DNN-2 more appropriately by comparing the restoration errors.

[Flowchart of DNN selection processing according to the first embodiment]
FIG. 7 is a diagram illustrating an example of a flowchart of the DNN selection process according to the first embodiment.

As shown in FIG. 7, the layer determination unit 15 and the restorer learning unit 11 select one DNN from a plurality of DNNs (step S11). The layer determination unit 15 determines from which intermediate layer of the selected DNN to take the intermediate feature amount (step S12).

The restorer learning unit 11 learns the restorer corresponding to the selected DNN with the general data (data group 21) by using the determined intermediate feature amount of the layer (step S13). For example, the restorer learning unit 11 inputs the data included in the data group 21 to the selected DNN. The restorer learning unit 11 extracts the intermediate feature amount obtained from the intermediate layer of the selected DNN, and inputs the extracted intermediate feature amount to the restorer corresponding to the selected DNN. The restorer learning unit 11 compares the restored data 22 output from the restorer with the data input to the selected DNN, and learns the restorer so that the error becomes small.

The layer determination unit 15 and the restorer learning unit 11 determine whether all DNNs have been selected (step S14). When it is determined that all the DNNs have not been selected (step S14; No), the layer determination unit 15 and the restorer learning unit 11 move to step S11 to select the next DNN.

On the other hand, when it is determined that all DNNs have been selected (step S14; Yes), the restored data generation unit 12 selects one DNN from the plurality of DNNs (step S15).

The restored data generation unit 12 inputs new data to be discriminated into the selected DNN, and the DNN selection unit 13 calculates the restored error (step S16). For example, the restored data generation unit 12 inputs new data to the selected DNN. The restored data generation unit 12 extracts the intermediate feature amount obtained from the intermediate layer of the selected DNN, and inputs the extracted intermediate feature amount to the restorer corresponding to the selected DNN. The restored data generator 12 outputs the restored data 22 from the restorer. The DNN selection unit 13 compares the restored data 22 output from the restorer with the new data, and calculates the restoration error.

The restored data generation unit 12 determines whether all DNNs have been selected (step S17). When it is determined that all the DNNs have not been selected (step S17; No), the restored data generation unit 12 proceeds to step S15 to select the next DNN.

On the other hand, when it is determined that all the DNNs have been selected (step S17; Yes), the DNN selection unit 13 selects the DNN corresponding to the restorer with the smallest restoration error (step S18). Then, the discrimination unit 14 inputs new data to be discriminated into the selected DNN and outputs the discrimination result (step S19).

[Effect of Example 1]
In this manner, the information processing device 1 uses, for each of the plurality of DNNs, a restorer that restores the data to be input to the DNN from the characteristics obtained from the intermediate layer of the DNN and the data used for learning of the plurality of DNNs. Learn with different data. The information processing device 1 selects, for new data, the DNN corresponding to the data output from each of the plurality of restorers and the restorer selected based on the new data. According to this configuration, the information processing device 1 can select a more appropriate DNN by using the data restored from the characteristics obtained from the intermediate layer of the DNN and the new data.

Further, the information processing device 1 calculates the error between the data output from each of the plurality of restorers and the new data, and selects the DNN corresponding to the restorer of the data having the smallest calculated error. According to such a configuration, the information processing device 1 can select a more appropriate DNN by using the error between the data output from the decompressor and the new data.

Further, the information processing device 1 sets the layer used as the intermediate layer to be a layer close to the fully connected layer when the DNN is the convolutional DNN. According to such a configuration, the information processing apparatus 1 can estimate the amount of information of the characteristics of the task itself handled by the DNN by setting the intermediate layer that obtains the characteristics of the DNN as a layer close to the fully connected layer.

In addition, the information processing apparatus 1 sets the layer used as the intermediate layer as the layer for the restorer that maximizes the variance of the restore error for the data used for learning for each restorer when each layer of the DNN is used. To do. According to this configuration, the information processing apparatus 1 has a case in which restoration can be performed successfully and a case in which restoration cannot be performed well if the variance of the restoration error is large. You can

By the way, in the first embodiment, the information processing apparatus 1 selects which DNN is to be applied to new data to be discriminated when there is no difference in the scale of the plurality of DNNs and the restorer. That is, the information processing device 1 learns, for each of the plurality of DNNs, a restorer that restores the data input to the DNN from the characteristics obtained from the intermediate layer (deep layer) of the DNN. Then, it is described that the information processing apparatus 1 selects the DNN corresponding to the data output from each of the plurality of restorers and the restorer selected based on the new data for the new data to be determined. However, the information processing apparatus 1 is not limited to this, and when there is a difference in the scale of a plurality of DNNs or a restorer, it is possible to select which DNN to apply to new data to be determined. Is also good. The scale of the DNN here means, for example, the number of layers and the number of units belonging to each layer.

Therefore, in the second embodiment, the information processing apparatus 1 selects which DNN to apply to new data to be determined when the plurality of DNNs and the size of the restorer are different. Will be described.

[Configuration of Information Processing Device According to Second Embodiment]
FIG. 8 is a functional block diagram of the configuration of the information processing apparatus according to the second embodiment. The same components as those of the information processing device 1 shown in FIG. The difference between the first embodiment and the second embodiment is that an auxiliary DNN learning unit 31, a restoration easiness calculation unit 32, and a certainty factor calculation unit 33 are added to the control unit 10. In addition, the data group 21 of the storage unit 20 is changed to a data group 21A. The point is that the DNN selection unit 13 of the control unit 10 is changed to the DNN selection unit 13A. The point is that class-based restoration ease 41 and class-based certainty factor 42 are added to the storage unit 20.

The data group 21A is a set of learning data such as a general natural image used for learning a restorer and an auxiliary DNN described later, and is a set of learning data with a label (class). The data group 21A is a set of learning data different from the learning data of each DNN. The data group 21A preferably includes various data such as ImageNet (150k images, 1000classes) and has a large number of classes.

The class-based restoration ease 41 is information indicating the ease of restoration for each class of the restorer, and is present for each restorer. The easiness of restoration of one class is, for example, the reciprocal of the average value of the restoration error between a plurality of data of the class included in the data group 21A and the restored data output from the restorer of these data. is there. That is, if the average value of the restoration error is small, the ease of restoration becomes large. If the average value of the restoration error is large, the ease of restoration is small. The class-based restoration ease 41 is generated by the restoration ease calculation unit 32.

The class-wise certainty factor 42 is information indicating the certainty factor for each class, which is output when the class is discriminated by the auxiliary DNN for new data. The auxiliary DNN referred to here is a learning device that discriminates a data class (label type). When learning, the auxiliary DNN learns using the same data as the data included in the data group 21A used for learning the restorer. The learned auxiliary DNN outputs the class-specific confidence 42, which is the confidence for each class, with respect to the new data.

The auxiliary DNN learning unit 31 learns the auxiliary DNN for discriminating the data class (label type). For example, the auxiliary DNN learning unit 31 learns with the same data as the data included in the data group 21 used for learning the restorer.

The easiness of restoration calculation unit 32 uses the data group 21A to calculate easiness of restoration for each class of the plurality of learned restorers. For example, the restoration easiness calculation unit 32 sequentially inputs the plurality of data included in the data group 21A to the plurality of DNNs. Hereinafter, the processing will be described by focusing on a specific DNN among a plurality of DNNs and a restorer corresponding to the specific DNN. The restoration easiness calculation unit 32 extracts, from the input data, a feature obtained from the middle layer of a specific DNN, and inputs the extracted feature to the restorer corresponding to the DNN. The restoration easiness calculation unit 32 compares the restoration data 22 output from the restoration device with the data input to the specific DNN to calculate the restoration error. The restoration easiness calculation unit 32 holds the restoration error calculated in association with the label (class) of the data input to the specific DNN. The restoration easiness calculation unit 32 holds the restoration error calculated in association with the label (class) of the data input to the specific DNN for the specific DNN even for other data of the plurality of data. .. Then, the restoration easiness calculation unit 32 calculates the reciprocal 41 of the restoration by class by calculating the reciprocal of the average value of the restoration error for each class for the restoration device corresponding to the specific DNN. The restoration easiness calculation unit 32 similarly calculates the restoration easiness 41 for each class for the restoration devices corresponding to other DNNs among the plurality of DNNs.

The certainty factor calculation unit 33 calculates the class certainty factor 42 for the new data to be discriminated through the learned auxiliary DNN.

The DNN selection unit 13A compares the class-based restoration ease 41 and the class-based certainty degree 42 of each of the plurality of restorers with respect to the new data, and the restorer having the class-based restoration ease 41 having the most similar tendency. And select the DNN corresponding to the selected restorer. For the similarity between the class-based restoration ease 41 and the class-based certainty factor 42, for example, the cosine similarity may be used. The ease of restoration by class 41 means information on the ease of restoration relative to the data group 21A in the DNN and the restorer. Therefore, the DNN selection unit 13A uses the ease of restoration 41 for each class to determine which DNN to use for new data to be discriminated, even when there is a difference in the scale of the plurality of DNNs and the restoration unit. You can choose whether to apply.

Here, an example of the flow of the restorer learning by the restorer learning unit 11 and the auxiliary DNN learning by the auxiliary DNN learning unit 31 will be described with reference to FIG. 9. FIG. 9 is a diagram illustrating an example of the flow of the restorer learning and the auxiliary DNN learning according to the second embodiment. The flow of the restorer learning by the restorer learning unit 11 has been described with reference to FIG.

The general data shown in FIG. 9 is an example of the data group 21A. The auxiliary DNN learning unit 31 inputs, to the auxiliary DNN, each labeled data included in the data group 21A used for learning the restorer. Then, the auxiliary DNN learning unit 31 learns in accordance with the input of the label of the input labeled data so that the determination result of the label of the input data approaches the label attached to the input data.

Here, an example of the flow of DNN selection by the DNN selection unit 13A will be described with reference to FIG. FIG. 10 is a diagram illustrating an example of the flow of DNN selection according to the second embodiment. FIG. 10 shows a plurality of DNNs and a learned restorer and an auxiliary DNN corresponding to each of the plurality of DNNs. In addition, in FIG. 10, it is assumed that the ease of restoration 41-1, 41-2, and 41-3 for each class of the learned restorers are calculated by the ease of restoration calculating unit 32.

Under such a situation, the certainty factor calculation unit 33 calculates the certainty factor by class 42 for the new data to be determined through the learned auxiliary DNN. Then, the DNN selection unit 13A compares the calculated class-based certainty factor 42 with the class-based restoration easiness 41-1 of the restorer-1. The DNN selection unit 13A compares the calculated class-based certainty factor 42 with the class-based restoration easiness 41-2 of the restoration unit-2. The DNN selecting unit 13A compares the calculated class-based certainty factor 42 with the class-based restoration ease 41-3 of the restoration unit-3. Then, as a result of the comparison, the DNN selecting unit 13A selects the restorer having the ease of class-based restoration 41 having the most similar tendency. Here, the restorer-2 is selected. As a result, the DNN selection unit 13A selects the DNN-2 corresponding to the selected restorer-2. That is, the DNN selection unit 13A applies DNN-2 because the class in which the DNN-2 is easily restored is similar to the class in which the certainty factor of new data is high.

As a result, the DNN selection unit 13A can select a more appropriate DNN for new data even when there is a difference in the scale of the plurality of DNNs and the restorer.

[Flowchart of DNN selection processing according to the second embodiment]
FIG. 11 is a diagram illustrating an example of a flowchart of the DNN selection process according to the second embodiment.

As shown in FIG. 11, the layer determination unit 15 and the restorer learning unit 11 select one DNN from the plurality of DNNs (step S21). The layer determination unit 15 determines from which intermediate layer of the selected DNN to take the intermediate feature amount (step S22).

The restorer learning unit 11 learns the restorer corresponding to the selected DNN with the general data (data group 21A) using the determined intermediate feature amount of the layer (step S23). For example, the restorer learning unit 11 inputs the data included in the data group 21A to the selected DNN. The restorer learning unit 11 extracts the intermediate feature amount obtained from the intermediate layer of the selected DNN, and inputs the extracted intermediate feature amount to the restorer corresponding to the selected DNN. The restorer learning unit 11 compares the restored data 22 output from the restorer with the data input to the selected DNN, and learns the restorer so that the error becomes small.

The layer determination unit 15 and the restorer learning unit 11 determine whether all DNNs have been selected (step S24). When it is determined that all the DNNs have not been selected (step S24; No), the layer determination unit 15 and the restorer learning unit 11 move to step S21 to select the next DNN.

On the other hand, when it is determined that all the DNNs have been selected (step S24; Yes), the auxiliary DNN learning unit 31 learns the auxiliary DNN using the labeled general data (data group 21A) (step S25).

Then, the easiness of restoration calculation unit 32 selects one DNN from the plurality of DNNs (step S26). The restoration easiness calculation unit 32 passes the labeled general data through the DNN and the restoration device, and calculates the restoration easiness for each class (step S27). For example, the restoration easiness calculation unit 32 calculates the reversion easiness 41 for each class by calculating the reciprocal of the average value of the restoration error for each class for the restoration device corresponding to the selected DNN.

The restoration easiness calculation unit 32 determines whether all DNNs have been selected (step S28). When it is determined that all the DNNs have not been selected (step S28; No), the restoration easiness calculation unit 32 proceeds to step S26 to select the next DNN.

On the other hand, when it is determined that all DNNs have been selected (step S28; Yes), the certainty factor calculation unit 33 inputs new data to be determined to the auxiliary DNN and calculates the certainty factor for each class. (Step S29). For example, the certainty factor calculation unit 33 calculates the certainty factor by class 42 for the new data to be determined through the learned auxiliary DNN.

Subsequently, the DNN selection unit 13A selects one DNN from the plurality of DNNs (step S30). The DNN selection unit 13A calculates the similarity between the restoration easiness for each class of the selected DNN (class-based restoration easiness 41) and the calculated certainty factor (class certainty factor 42) ( Step S31). The similarity is, for example, the cosine similarity.

The DNN selection unit 13A determines whether or not all DNNs have been selected (step S32). When it is determined that all the DNNs are not selected (step S32; No), the DNN selection unit 13A proceeds to step S30 to select the next DNN.

On the other hand, when it is determined that all DNNs have been selected (step S32; Yes), the DNN selection unit 13A selects the DNN with the highest degree of similarity (step S33). For example, the DNN selecting unit 13A selects a restorer having the ease of restoration 41 for each class, which has the most similar tendency, and selects the DNN corresponding to the selected restorer. Then, the discrimination unit 14 inputs new data to be discriminated into the selected DNN and outputs the discrimination result (step S34).

[Effect of Embodiment 2]
In this way, the information processing apparatus 1 uses, for each of the plurality of DNNs, a restorer that restores the data input to the DNN from the characteristics obtained from the intermediate layer of the DNN, as well as the data used for learning of each of the plurality of DNNs. Learn with different data. The information processing device 1 learns the auxiliary DNN for discriminating the class of data by the same data as the data used for learning the restorer. The information processing apparatus 1 outputs information indicating the ease of restoration for each class of a plurality of restorers (restorability for each class 41) and the class output when new data is learned by the auxiliary DNN. The certainty factor (class certainty factor 42) is compared. Then, as a result of the comparison, the information processing device 1 selects the DNN corresponding to the restorer of the class-based restoration ease 41 having the most similar tendency. With this configuration, the information processing apparatus 1 can select a more suitable DNN for new data even when there is a difference in the scale of the plurality of DNNs and the restorer.

[Other]
The components of the illustrated information processing apparatus 1 do not necessarily have to be physically configured as illustrated. That is, the specific aspect of the distribution/integration of the information processing device 1 is not limited to that shown in the figure, and all or part of the information processing device 1 may be functionally or physically arranged in arbitrary units according to various loads and usage conditions. It can be distributed and integrated. For example, the restoration data generation unit 12 and the DNN selection unit 13 may be integrated as one unit. Further, the layer determining unit 15 is separated into a first layer determining unit that determines a layer before the fully connected layer as an intermediate layer and a second layer determining unit that determines the intermediate layer by using the variance of the restoration error. You may. Further, the storage unit 20 may be connected as an external device of the information processing device 1 via a network.

The various processes described in the first and second embodiments can be realized by executing a prepared program on a computer such as a personal computer or a workstation. Therefore, an example of a computer that executes a DNN selection program that realizes the same function as the information processing device 1 illustrated in FIG. 1 will be described below. FIG. 12 is a diagram illustrating an example of a computer that executes a DNN selection program.

As shown in FIG. 12, the computer 200 includes a CPU 203 that executes various arithmetic processes, an input device 215 that receives data input from a user, and a display control unit 207 that controls the display device 209. The computer 200 also includes a drive device 213 that reads a program and the like from a storage medium, and a communication control unit 217 that exchanges data with another computer via a network. The computer 200 also has a memory 201 for temporarily storing various information and an HDD (Hard Disk Drive) 205. The memory 201, the CPU 203, the HDD 205, the display control unit 207, the drive device 213, the input device 215, and the communication control unit 217 are connected by the bus 219.

The drive device 213 is, for example, a device for the removable disk 210. The HDD 205 stores a DNN selection program 205a and DNN selection processing related information 205b.

The CPU 203 reads the DNN selection program 205a, loads it into the memory 201, and executes it as a process. Such a process corresponds to each functional unit of the information processing device 1. The DNN selection processing related information 205b corresponds to the data group 21, the restored data 22, and the restored data (new data) 23. Then, for example, the removable disk 210 stores each information such as the DNN selection program 205a.

The DNN selection program 205a does not necessarily have to be stored in the HDD 205 from the beginning. For example, a "portable physical medium such as a flexible disk (FD), a CD-ROM (Compact Disk Read Only Memory), a DVD (Digital Versatile Disk), a magneto-optical disk, an IC (Integrated Circuit) card, which is inserted into the computer 200. The relevant program is stored in. Then, the computer 200 may read the DNN selection program 205a from these and execute it.

DESCRIPTION OF SYMBOLS 1 Information processing apparatus 10 Control part 11 Restoration learning part 12 Restoration data generation part 13, 13A DNN selection part 14 Discrimination part 15 Layer determination part 31 Auxiliary DNN learning part 32 Restorability calculation part 33 Certainty factor calculation part 20 Storage part 21,21A Data group 22 Restored data 23 Restored data (new data)
41 Ease of restoration by class 42 Confidence by class

Claims (7)

On the computer,
For each of a plurality of deep learning neural networks (DNNs), a restorer that restores the data input to the DNN from the characteristics obtained from the middle layer of the DNN is learned by different data from the data used for the learning of each of the plurality of DNNs. Then
A DNN selection program for executing a process of selecting, for specific data, data output from each of a plurality of restorers and a DNN corresponding to the restorer selected based on the specific data. The selecting process calculates an error between the data output from each of the plurality of decompressors and the specific data, and selects the DNN corresponding to the decompressor of the data having the smallest calculated error. The DNN selection program according to claim 1. A learning device that discriminates a class of data is learned by the same data as the data used for learning the restorer,
The process of selecting includes information indicating the ease of restoration for each class of the plurality of restorers, and confidence for each class output when the specific data is learned by the learning device. The DNN selection program according to claim 1, wherein the DNN corresponding to the restorer of the information having the most similar tendency is selected. The DNN selection program according to claim 1, wherein the layer used as the intermediate layer is a layer close to a fully connected layer when the DNN is a convolutional DNN. The layer used as the intermediate layer is a layer for the restorer that maximizes the variance of the restore error with respect to the data used for learning for each restorer when each layer of the DNN is used. The DNN selection program according to claim 1. For each of a plurality of deep learning neural networks (DNNs), a restorer that restores the data input to the DNN from the characteristics obtained from the middle layer of the DNN is learned by different data from the data used for the learning of each of the plurality of DNNs. Then
A DNN selection method in which a computer executes a process of selecting, for specific data, data output from each of a plurality of restorers and a DNN corresponding to the restorer selected based on the specific data. For each of a plurality of deep learning neural networks (DNNs), a restorer that restores the data input to the DNN from the characteristics obtained from the middle layer of the DNN is learned by different data from the data used for the learning of each of the plurality of DNNs. A restorer learning unit that
A DNN selecting unit that selects the DNN corresponding to the data output from each of the plurality of restorers learned by the restorer learning unit and the restorer selected based on the specific data for the specific data;
An information processing device comprising:

Images