KR20210038027A - Method for Training to Compress Neural Network and Method for Using Compressed Neural Network - Google Patents

Abstract

Disclosed are a neural network compression training method and a method for using a compressed neural network. According to one aspect of the present invention, provided is a method for compressing a neural network, and more particularly, a neural network compression method for reducing the number of parameters to be trained by a neural network to reduce the computational throughput and computational speed of the neural network. The neural network compression training method comprises the following processes of: obtaining a plurality of batches and providing the plurality of batches to a neural network; and updating parameters of the neural network for each batch.

Description

Neural Network Compression Training Method and Method for Using Compressed Neural Network {Method for Training to Compress Neural Network and Method for Using Compressed Neural Network}

Embodiments of the present invention relate to a method of compressing a neural network, in particular, to a method of compressing a neural network in which the number of parameters to be learned by the neural network is reduced to reduce the computational throughput and the computation speed of the neural network.

The content described in this section merely provides background information on the present invention and does not constitute prior art.

Recently, an image classification method or image style transfer method based on deep learning uses a neural network having a number of channels. Such a neural network needs to learn a large number of parameters to perform a general image classification task or an image style conversion method. There is a problem that as the number of parameters that the neural network needs to learn increases, the computational throughput and power consumption of the processor increase exponentially. Therefore, it is necessary to compress the neural network in terms of processing throughput, processing speed, and power consumption of the processor.

A channel pruning method was devised to compress neural networks. The channel pruning method is a method of training a neural network to remove a small parameter (or weight) from a neural network after learning a parameter until a loss function of a neural network decreases to a specific criterion.

In particular, the conventional channel pruning method used in a convolution neural network is performed in two steps. The first step is to compress the neural network by removing the small filter parameters from the trained neural network, and the second step is to retrain the filter parameters of the compressed neural network.

However, the existing channel pruning method removes a parameter having a small size by considering only the size of a filter parameter without considering the size of an element included in an input feature map input to a convolutional neural network. In a convolutional neural network, even if the size of the filter parameter is small, when the size of an element included in the input feature map is large, the filter may generate a feature map including the element having a large size. As described above, if all filter parameters having a small size are removed irrespective of the size of the input feature map, the output feature map with a large size that can be generated by the small filter parameter is also removed, thereby reducing the performance of the neural network.

In addition, in the conventional channel pruning method, after performing the step of compressing the neural network, it is necessary to additionally perform the step of retraining the filter parameters of the compressed neural network. This is a process for compensating for performance degradation caused by the step of removing the filter parameter. Accordingly, the conventional channel pruning method has a problem in that the time required to train the neural network increases because the process of compressing and retraining the neural network is required.

In embodiments of the present invention, the number of parameters to be learned by the neural network is determined by removing a filter that generates a zero feature map based on a feature map generated by the filter regardless of the size of the input feature map input to the filter in the neural network. The main purpose is to reduce.

Another embodiment of the present invention maintains the performance of a neural network by removing the filter that generates the zero feature map after training to output all zero feature maps or all thesis feature maps for an arbitrary image. At the same time, it aims to reduce the number of learning parameters.

According to an aspect of the present invention, in a training method for compressing a neural network that generates feature maps using a plurality of filters for a plurality of batches including a plurality of images, the plurality of batches are obtained, and the plurality of Providing an arrangement to the neural network; And updating parameters of the neural network for each batch, wherein the updating includes: obtaining a plurality of feature maps corresponding to the images included in the batch using the neural network; Dividing the plurality of feature maps corresponding to the images into a zero feature map and a non-zero feature map; And modifying a parameter of the neural network using the number of zero feature maps and the number of filters that output a zero feature map or a topic zero feature map for two different images among the plurality of images, The zero feature map is a feature map in which all elements are 0, and the non-zero feature map provides a training method in which at least one of the elements is a non-zero feature map.

According to another aspect of the present embodiment, there is provided an image processing method, comprising: obtaining an input image; Providing the input image to a first neural network, wherein the first neural network generates feature maps for the input image using a plurality of filters; And obtaining, by a second neural network, a result of processing the image using the feature maps received from the first neural network, wherein the first neural network It is trained with two neural networks, and the training method includes: acquiring a plurality of feature maps corresponding to the images included in the batch using the first neural network for each batch; Dividing the plurality of feature maps corresponding to each image into a zero feature map and a non-zero feature map; And modifying parameters of the first neural network using the number of zero feature maps and the number of filters that output zero feature maps or non-zero feature maps for two different images among the plurality of images. However, the zero feature map is a feature map in which all elements are 0, and the non-zero feature map provides an image processing method in which at least one of the elements is a non-zero feature map.

As described above, according to an embodiment of the present invention, by removing a filter that generates a zero feature map based on a feature map generated by the filter regardless of the size of the input feature map input to the filter in the neural network, the neural network By reducing the number of parameters to be learned, the computational throughput and computational speed of the neural network can be reduced.

According to another embodiment of the present invention, the performance of the neural network is improved by removing the filter that generates the zero feature map after training to output all zero feature maps or all thesis feature maps for an arbitrary image. While maintaining, it is possible to reduce the number of learning parameters to reduce the computational throughput and computational speed of the neural network.

1A and 1B are diagrams for explaining a channel pruning method among methods of compressing a neural network in a convolutional neural network.
2 is a diagram illustrating a process of compressing a neural network according to an embodiment of the present invention.
3 is a diagram for explaining a neural network compression method according to an embodiment of the present invention in terms of equations.
4 is a flowchart illustrating a compression training method of a neural network according to an embodiment of the present invention.
5 is a flowchart illustrating a method of using a compressed neural network according to an embodiment of the present invention.

Hereinafter, some embodiments of the present invention will be described in detail through exemplary drawings. In adding reference numerals to elements of each drawing, it should be noted that the same elements are assigned the same numerals as possible, even if they are indicated on different drawings. In addition, in describing the present invention, when it is determined that a detailed description of a related known configuration or function may obscure the subject matter of the present invention, a detailed description thereof will be omitted.

In addition, in describing the constituent elements of the present invention, terms such as first, second, A, B, (a) and (b) may be used. These terms are for distinguishing the constituent element from other constituent elements, and the nature, order, or order of the constituent element is not limited by the term. Throughout the specification, when a part'includes' or'includes' a certain element, it means that other elements may be further included rather than excluding other elements unless otherwise stated. . In addition, terms such as'~ unit' and'module' described in the specification mean a unit that processes at least one function or operation, which may be implemented as hardware or software, or a combination of hardware and software.

1A and 1B are diagrams for explaining a channel pruning method among methods of compressing a neural network in a convolutional neural network.

1A, an input feature map 100, a kernel, that is, a first filter 102, an intermediate feature map 104, a second filter 106, and an output feature map 108 are It is illustrated.

The input feature map 100 is extracted from the input image and will be described as having three channels. Three input feature maps 100 are input to five first filters 102. Each filter creates one intermediate feature map 104 for three channels. That is, the number of intermediate feature maps 104 is determined by the number of first filters. Five intermediate feature maps 104 are input to two second filters 106, and each filter generates two output feature maps 108 for five intermediate feature maps 104.

In order to apply the channel pruning method, a filter with a small parameter size included in the filter is removed. When the size of the parameters included in the two filters among the five first filters 102 is small, it is determined that the two filters are filters that do not affect the performance of the neural network, and the two intermediate filters are generated by the two filters. The feature map can also be determined as a feature map that does not affect the performance of the neural network.

Referring to FIG. 1B, a process of re-training a neural network after removing two filters composed of small parameters is shown.

The three input feature maps 100 are input to the three first filters 112, and the three first filters 112 generate three intermediate feature maps 114. Three intermediate feature maps 114 are input to two second filters 116, and two second filters 116 generate two output feature maps 118.

At this time, before the two filters are removed, two intermediate feature maps generated by the two filters are also removed together with the two filters. In this case, even if the size of the parameters included in the two filters is small, the two intermediate feature maps may include a parameter having a large size. That is, by removing the two filters, even a feature map that may affect the performance of the neural network is removed, so that the performance of the neural network may be degraded.

In order to prevent the performance of the neural network from deteriorating due to the removal of two of the five input filters, the neural network performs a process of retraining the parameters included in the three first filters 112.

2 is a diagram illustrating a process of compressing a neural network according to an embodiment of the present invention.

Referring to FIG. 2, a first image 210, a second image 220, a neural network 230, four filters 232, 234, 236, 238, and four first feature maps for the first image. (212, 214, 216, 218) and four second feature maps 222, 224, 226, 228 for the second image are shown.

The first image 210 and the second image 220 are training data of a neural network. However, the first image 210 and the second image 220 are only one embodiment, and are not limited thereto, and may be data representing a plurality of characters in a multidimensional matrix.

The neural network 230 includes four filters 232, 234, 236, and 238, but this is only one embodiment and may include a plurality of filters. The neural network 230 receives the first image 210 and the second image 220, and generates feature maps using four filters 232, 234, 236, and 238. The first feature maps 212, 214, 216, and 218 and the second feature maps 222, 224, 226, and 228 include a plurality of elements for each feature map.

The first feature maps 212, 214, 216, 218 are feature maps generated by the neural network 230 for the first image 210, and the second feature maps 222, 224, 226, 228 are neural networks. Reference numeral 230 denotes feature maps generated for the second image 220. Specifically, the first filter 232 generates a feature map 212 for the first image and a feature map 222 for the second image. Likewise, the second filter 234, the third filter 236, and the fourth filter 238 respectively include feature maps 214, 216, and 218 for the first image and feature maps 224 for the second image, respectively. , 226, 228).

The first feature maps 212, 214, 216, 218 and the second feature maps 222, 224, 226, 228 are a zero feature map and a non-zero feature map. It is divided into. Here, the zero feature map means a feature map in which values of all elements are 0, and the non-zero feature map means a feature map in which at least one of the elements is not zero. Some of the first feature maps 212 and 214 are topic-zero feature maps, and the rest of the first feature maps 216 and 218 are zero feature maps. Some of the second feature maps 214 and 216 are topic-zero feature maps, and the rest of the second feature maps 212 and 218 are zero feature maps.

The parameters of the neural network 230 include the number of zero feature maps 216, 218, 222, 228, and the filter 238 outputting a zero feature map for both the first image 210 and the second image 220. May be modified based in part on the number.

In the training method for compressing a neural network according to an embodiment of the present invention, the number of filters for generating zero feature maps for both the first image 210 and the second image 220 is increased. The parameters of the neural network 230 are modified.

A training method for compressing a neural network according to another embodiment of the present invention includes a parameter of the neural network 230 so that the number of zero feature maps among a plurality of feature maps generated for the first image 210 and the second image 220 is increased. Modify it.

In the training method for compressing a neural network according to an embodiment of the present invention, parameters of the neural network may be updated in units of each batch for a plurality of batches. Here, the arrangement is a unit including a plurality of images that are training data. After updating the parameters of the neural network for a plurality of batches, two different images among the plurality of images are provided to the neural network. Among the plurality of filters included in the neural network, filters that generate zero feature maps for two different images are removed.

Specifically, in the training method for compressing a neural network, one arrangement among a plurality of arrangements is provided to the neural network, and a plurality of feature maps corresponding to each image included in one arrangement is obtained using the neural network. A plurality of feature maps corresponding to each image is divided into a zero feature map and a topic zero feature map. The parameters of the neural network may be modified based on the number of filters that output a zero feature map or a non-zero feature map for two different images included in one arrangement and the number of zero feature maps. Thereafter, the neural network can be trained by repeating the above process for each batch. For example, when two batches are given, and each batch includes eight images, the neural network training method may update the parameters of the neural network 28 times for one batch. No. 28 is calculated from C(8, 2), and C(8, 2) means a combination of two different images among the eight images.

The compressed neural network can be trained through the processes described in FIG. 2, and since training and compression for the parameters of the neural network are performed at the same time, the number of parameters that the neural network learns is reduced, but a separate retraining process is not required and Performance is maintained.

3 is a diagram for explaining a neural network compression method according to an embodiment of the present invention in terms of equations.

The neural network compression training method according to an embodiment of the present invention uses a loss function including two losses. The two losses are channel loss and xor loss (exclusive or loss).

The channel loss is a parameter that increases the number of zero feature maps among a plurality of feature maps output from the convolutional layer. That is, as the channel loss decreases, the number of zero feature maps among the output feature maps increases.

The xor loss is a parameter for removing a filter that generates a zero feature map even if the input image is different. Using xor loss, it is possible to generate zero feature maps for both input images or increase the number of filters that generate non-zero feature maps for both. That is, as the xor loss decreases, the number of filters that always output zero feature maps for any input image or all non-zero feature maps increases.

According to an embodiment of the present invention, by removing a filter that always generates a zero feature map for an arbitrary image using two losses, it is possible to maintain the performance of the neural network and reduce the number of training parameters of the neural network.

은 L번째 컨볼루션 레이어로부터 출력되는 특징맵들이고,

은 L번째 레이어로부터 출력되는 특징맵들의 개수를 나타내는 채널이다. Referring to FIG. 3, B is a unit for updating a parameter of a neural network, and a convolution layer is included in the neural network, includes a plurality of filters, and outputs a plurality of feature maps.

Are feature maps output from the L-th convolution layer,

Is a channel indicating the number of feature maps output from the L-th layer.

와 이미지 j에 대한 복수의 특징맵인

에 대해 L0-정규화(L0-normalization)를 통해 제로 특징맵과 논제로 특징맵으로 구분한다. xor 로스는 이미지 i 및 이미지 j에 대해 모두 제로 특징맵을 생성하거나 모두 논제로 특징맵을 생성하는 필터의 수를 합한 값에 부분적으로 기초하여 계산된다. 채널 로스는 이미지 i 및 이미지 j에 대해 논제로 특징맵을 생성하는 필터의 수를 모두 합한 값에 부분적으로 기초하여 계산된다.In the training method according to an embodiment of the present invention, image i and image j included in batch B are provided to a neural network, and feature maps are obtained from convolutional layer l included in the neural network. After that, a plurality of feature maps for image i

And multiple feature maps for image j

Is divided into a zero feature map and a topic zero feature map through L0-normalization. The xor loss is calculated based in part on the sum of the number of filters that generate zero feature maps for both images i and j, or all non-zero feature maps. The channel loss is calculated based in part on the sum of all the number of filters that generate the feature map as a topic for image i and image j.

는 L0-norm으로서, 특징맵 x에 포함된 모든 엘리먼트가 0이면 0을 반환하고, 엘리먼트 중 적어도 하나가 0이 아니면 1을 반환하는 기호이다. 또한, 수학식 2에서 t는 채널의 인덱스(index)를 의미한다.Specifically, the channel loss and xor loss are calculated by Equations 1 and 2, respectively. In Equation 1 and Equation 2

Is L0-norm, and is a symbol that returns 0 if all elements included in the feature map x are 0, and returns 1 if at least one of the elements is not 0. In addition, in Equation 2, t denotes an index of a channel.

는 수학식 3과 같이 수정될 수 있다. 수학식 3에서

는 L2-norm(L2 정규화)이다. 이는, 채널 로스 및 xor 로스를 포함하는 손실함수를 미분가능하도록 하기 위함이다.

일 수 있다.According to an embodiment of the present invention.

Can be modified as in Equation 3. In Equation 3

Is L2-norm (L2 normalized). This is to make it possible to differentiate a loss function including a channel loss and an xor loss.

Can be

4 is a flowchart illustrating a compression training method of a neural network according to an embodiment of the present invention.

The processes described below may be performed using one or more memories and one or more processors.

First, a plurality of images is acquired from a memory storing a plurality of images (S400). In this case, an image may be acquired in a batch unit including a plurality of images.

A plurality of arrangements are provided to the neural network, and the neural network receives a plurality of feature maps generated by using a plurality of filters for each image included in one arrangement (S402).

A plurality of feature maps corresponding to each image is divided into a zero feature map and a topic zero feature map (S404).

The parameters of the neural network are modified to increase the number of filters that generate zero feature maps or all non-zero feature maps for two different images among the plurality of images (S406).

The parameters of the neural network are modified to increase the number of zero feature maps among the plurality of feature maps (S408).

In the neural network training method according to an embodiment of the present invention, steps S402 to S408 may be repeatedly performed for each batch. In addition, the neural network training method may simultaneously perform steps S406 and S408.

5 is a flowchart illustrating a method of using a compressed neural network according to an embodiment of the present invention.

As a method of processing an image using a compressed neural network, a method of using the trained first neural network and the trained second neural network will be described.

Referring to FIG. 5, a plurality of input images are acquired from a memory (S500).

The obtained input images are provided to the first neural network (S502). The first neural network generates feature maps for each of a plurality of input images using a plurality of filters. Feature maps generated by the first neural network are transmitted to the second neural network.

The second neural network obtains a result of processing the image using the feature maps received from the first neural network (S504).

Here, when the second neural network is an image classifier, a result of classifying a plurality of images by the second neural network may be obtained.

On the other hand, when the second neural network is an image style converter, it is possible to obtain a composite image to which image style conversion is applied to two different images among a plurality of images. Image style conversion refers to a technique of extracting a content feature from one image of two images, extracting a style feature from the other image, and then outputting an image in which the content feature and the style feature are combined.

Meanwhile, the first neural network and the second neural network may be trained together. It will be described below.

In the training method, first, a plurality of feature maps corresponding to images included in the batch are obtained using a first neural network for each batch among a plurality of batches. The plurality of feature maps generated by the first neural network are transmitted to the second neural network. Thereafter, a plurality of feature maps corresponding to the images are divided into a zero feature map and a topic zero feature map. The parameters of the first neural network are modified using the number of zero feature maps and the number of filters that output zero feature maps or all non-zero feature maps for two different images among a plurality of images.

According to an embodiment of the present invention, when the first neural network and the second neural network are trained to classify a plurality of images, the second neural network may be an image classifier. In the training method, first, a plurality of images is provided to the first neural network. The first neural network generates a plurality of feature maps for a plurality of images and delivers them to the second neural network, and the second neural network labels a plurality of images using the plurality of feature maps. A loss function is extracted from a plurality of labeled images and preset correct answer data. The channel loss and xor loss of the first neural network are added to the extracted loss function. Finally, the first neural network and the second neural network may be trained by modifying the parameters of the first neural network and the parameters of the second neural network together so that the loss function including the channel loss and the xor loss is reduced to a preset value.

On the other hand, according to an embodiment of the present invention, when the first neural network and the second neural network are trained to convert the image style, the first neural network may be an encoder and the second neural network may be a decoder.

In the training method, first, two different images are provided to the first neural network. The first neural network extracts a feature map related to content from one image, and extracts a feature map related to style from the other image. The first neural network transmits the feature map related to the content and the feature map related to the style to the second neural network. The second neural network generates an image in which content and style are synthesized using a feature map related to the content and a feature map related to the style. The loss function is extracted from the synthesized image and the preset correct answer data. The channel loss and xor loss of the first neural network are added to the extracted loss function. Finally, the first neural network and the second neural network may be trained by modifying the parameters of the first neural network and the parameters of the second neural network together so that the loss function including the channel loss and the xor loss is reduced to a preset value.

In FIGS. 4 and 5, it is described that processes S400 to S504 are sequentially executed, but this is merely illustrative of the technical idea of an embodiment of the present invention. In other words, a person of ordinary skill in the art to which an embodiment of the present invention belongs can change the order shown in FIGS. 4 and 5 and execute the procedure S400 to the process without departing from the essential characteristics of the embodiment of the present invention. Since one or more processes in S504 are executed in parallel, various modifications and variations may be applied, and thus FIGS. 4 and 5 are not limited to a time-series order.

Meanwhile, the processes shown in FIGS. 4 and 5 can be implemented as computer-readable codes on a computer-readable recording medium. The computer-readable recording medium includes all types of recording devices that store data that can be read by a computer system. That is, the computer-readable recording medium includes storage media such as magnetic storage media (eg, ROM, floppy disk, hard disk, etc.) and optical reading media (eg, CD-ROM, DVD, etc.). In addition, the computer-readable recording medium can be distributed over a computer system connected through a network to store and execute computer-readable codes in a distributed manner. Further, it includes one or more processors and one or more memories for storing instructions, wherein when the instruction is executed by the one or more processors, the instruction causes the one or more processors to perform steps S400 to S408 or steps S500 to S504. Let's do it.

The above description is merely illustrative of the technical idea of the present embodiment, and those of ordinary skill in the technical field to which the present embodiment pertains will be able to make various modifications and variations without departing from the essential characteristics of the present embodiment. Accordingly, the present embodiments are not intended to limit the technical idea of the present embodiment, but to explain the technical idea, and the scope of the technical idea of the present embodiment is not limited by these embodiments. The scope of protection of this embodiment should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present embodiment.

100: input feature map 108: output feature map
210: first image 220: second image
230: neural network

Claims (16)

In a training method for compressing a neural network that generates feature maps using a plurality of filters for a plurality of batches including a plurality of images,
Obtaining the plurality of arrangements and providing the plurality of arrangements to the neural network; And
Including the process of updating the parameters of the neural network for each batch,
The updating process,
Obtaining a plurality of feature maps corresponding to the images included in the arrangement using the neural network;
Dividing the plurality of feature maps corresponding to the respective images into a zero feature map and a non-zero feature map; And
A process of modifying a parameter of the neural network using the number of zero feature maps and the number of filters that output a zero feature map or a topic zero feature map for two different images among the plurality of images,
The zero feature map is a feature map in which all elements are 0, and the non-zero feature map is a feature map in which at least one of the elements is non-zero. The method of claim 1,
The training method wherein the neural network includes at least one convolution layer, and the plurality of filters are included in the at least one convolution layer. The method of claim 1,
The modification process,
And modifying the parameter to increase the number of zero feature maps and the number of filters. As an image processing method,
Obtaining an input image;
Providing the input image to a first neural network, wherein the first neural network generates feature maps for the input image using a plurality of filters; And
A second neural network obtaining a result of processing the image using the feature maps received from the first neural network,
The first neural network is trained with the second neural network for a plurality of batches including a plurality of images, and the training method is for each batch.
Acquiring a plurality of feature maps corresponding to the images included in the arrangement using the first neural network;
Dividing the plurality of feature maps corresponding to each image into a zero feature map and a non-zero feature map; And
Modifying parameters of the first neural network using the number of zero feature maps and the number of filters outputting a zero feature map or a non-zero feature map for two different images among the plurality of images;
Including,
The zero feature map is a feature map in which all elements are 0, and the non-zero feature map is a feature map in which at least one of the elements is non-zero. The method of claim 4,
The first neural network includes at least one convolutional layer, and the plurality of filters are included in the at least one convolutional layer. The method of claim 4,
The modification process,
And modifying the parameter to increase the number of zero feature maps and the number of filters. The method of claim 4,
The first neural network is an encoder, and the second neural network is a decoder. The method of claim 4,
The first neural network is an encoder, and the second neural network is an image classifier. In a system for compressing a neural network,
One or more processors; And
Including one or more memories for storing instructions,
When the instruction is executed by the one or more processors, the instruction causes the one or more processors to perform a neural network training method,
The training method,
A training method for compressing the neural network generating feature maps using a plurality of filters for a plurality of batches including a plurality of images,
Acquiring the plurality of arrangements and providing the plurality of arrangements to the neural network; And
Including the process of updating the parameters of the neural network for each batch,
The updating process,
Obtaining a plurality of feature maps corresponding to the images included in the arrangement using the neural network;
Dividing the plurality of feature maps corresponding to each of the images into a zero feature map and a non-zero feature map; And
A process of modifying a parameter of the neural network using the number of zero feature maps and the number of filters that output a zero feature map or a topic zero feature map for two different images among the plurality of images,
The zero feature map is a feature map in which all elements are 0, and the non-zero feature map is a feature map in which at least one of the elements is not zero. The method of claim 9,
Wherein the neural network includes at least one convolutional layer, and the plurality of filters are included in the at least one convolutional layer. The method of claim 9,
The modification process,
And modifying the parameter to increase the number of zero feature maps and the number of filters. In a system for processing images,
One or more processors; And
Including one or more memories for storing instructions,
When the instruction is executed by the one or more processors, the instruction causes the one or more processors to perform an image processing method,
The image processing method,
Obtaining an input image;
Providing the input image to a first neural network, wherein the first neural network generates feature maps for the input image using a plurality of filters; And
A second neural network obtaining a result of processing the image using the feature maps received from the first neural network,
The first neural network is trained with the second neural network for a plurality of batches including a plurality of images, and the training method is for each batch.
Acquiring a plurality of feature maps corresponding to the images included in the arrangement using the first neural network;
Dividing the plurality of feature maps corresponding to each image into a zero feature map and a non-zero feature map; And
Modifying parameters of the first neural network using the number of zero feature maps and the number of filters outputting a zero feature map or a non-zero feature map for two different images among the plurality of images;
Including,
The zero feature map is a feature map in which all elements are 0, and the non-zero feature map is a feature map in which at least one of the elements is not zero. The method of claim 12,
Wherein the first neural network includes at least one convolutional layer, and the plurality of filters are included in the at least one convolutional layer. The method of claim 12,
The modification process,
And modifying the parameter to increase the number of zero feature maps and the number of filters. The method of claim 12,
The first neural network is an encoder, and the second neural network is a decoder. The method of claim 12,
The first neural network is an encoder, and the second neural network is an image classifier.

Images