Image classification method, device, medium and system for maximizing mutual information

Abstract

The present invention provides an image classification method, device, medium and system for maximizing mutual information, including: collecting training images; maximizing the mutual information between the training images and the neural network structure, automatically determining the network structure and parameters of the neural network; using the obtained neural network to process the image data to be classified to obtain the image classification result. Based on the given image data, the present invention automatically designs and determines the network structure and parameters of the neural network by maximizing the mutual information, which is used for image classification, without complicated manual design, saving human resources and computing resource consumption. The present invention can automatically design an image classification method based on a neural network in a very short time, and can achieve a high image classification accuracy rate at the same time.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and image processing technology, and in particular to an image classification method for maximizing mutual information, and a computer device and a readable storage medium thereof. Background Art

[0002] With the development of big data and the improvement of computing power, artificial intelligence technology has achieved rapid development in recent years, and the industry has increasingly demanded image processing applications such as image classification. Early image processing was limited by acquisition equipment, resulting in low image quality. The main technology used was manual feature extraction. Some manually designed feature extraction operators such as HoG and signal processing methods such as wavelet analysis promoted the early development of image processing. However, the manually designed feature extraction operators are designed by scientists based on their prior knowledge and analysis of a given image. They inevitably retain the bias brought by human priors, and their performance in image classification tasks cannot surpass that of humans. Neural networks, on the other hand, combine feature extractors and classifiers for end-to-end training. The big data-driven approach enables the network to automatically learn the feature extraction filter group that is most suitable for image classification targets. The emergence of neural networks has abandoned the manual design of feature extraction operators and has achieved performance that exceeds that of human experts in image classification tasks. Although the emergence of neural networks has eliminated the need for manually designed feature extraction operators, saving human resources and improving image classification performance, the architecture of the neural network itself still relies on manual design and construction.

[0003] In the past decade, hand-designed neural networks have achieved good performance in image processing tasks such as image classification. However, the increasingly complex neural networks and the diverse image classification requirements have made the design of neural networks a cumbersome, inefficient, and labor-intensive and computationally intensive task. The image classification method that maximizes mutual information gives machines the ability to autonomously design neural networks for specific image classifications, and can automatically design neural network models with better performance in a very short time, providing a more efficient and convenient solution for building neural networks for image classification tasks in industrial applications, such as efficient application to different types of image data (such as natural images, medical images, etc.), and efficient configuration on devices with different computing resources (such as central servers, edge devices, etc.). With the development of big data and the improvement of computing power, image classification methods based on neural networks have received certain attention and development in recent years. For example, the image classification accuracy of neural network architectures such as Googlenet, Resnet, and Densenet on the natural image dataset ImageNet has exceeded that of humans. However, the complex human resource consumption, lengthy network design time, and specific datasets make the practicality of image classification methods based on neural networks poor. Therefore, new neural network-based image classification methods are needed that can significantly reduce the time of neural network design while ensuring high performance so that image classification methods can be efficiently deployed in actual industrial applications.

[0004] In addition, the existing neural network-based image classification methods have the following shortcomings: (1) For specific image data, experts manually design neural network-based image classification methods. The design process is complicated and requires a lot of human resources and computing resources; (2) Due to the limitations of expert knowledge, the manually designed neural network-based image classification methods are not the best neural networks, and the image classification performance has a lot of room for improvement. (3) The existing automatic neural network design image classification methods have a high computational cost and take a lot of time. At the same time, the performance also has a certain room for improvement. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides an image classification method that maximizes mutual information. The method can automatically determine the network structure and parameters of a neural network based on given image data for image classification, greatly reducing design time and human resource consumption, while achieving a higher image classification accuracy.

[0006] According to one aspect of the present invention, there is provided an image classification method for maximizing mutual information, comprising:

[0007] Collect training images;

[0008] Maximize the mutual information between the training image and the neural network structure, and automatically determine the network structure and parameters of the neural network;

[0009] The obtained neural network is used to process the image data to be classified to obtain an image classification result.

[0010] Preferably, the collected training image is divided into two parts; the maximizing the mutual information between the training image and the neural network structure and automatically determining the network structure and parameters of the neural network includes:

[0011] Constructing a hypernetwork and a structure generation network, performing data processing on them to obtain the network parameters of the hypernetwork and the parameters of the structure generation network, and constructing a target network;

[0012] All training images are input into the target network to generate predicted image category labels, and the cross entropy loss of image classification is calculated based on the predicted image category labels and the true image category labels. The target network is trained until convergence for image classification.

[0013] Preferably, the step of constructing a hypernetwork and a structure generation network, performing data processing on them to obtain network parameters of the hypernetwork and parameters of the structure generation network, and constructing a target network comprises:

[0014] S1, constructs basic units based on all possible operations for image classification;

[0015] The basic unit is then used to construct a super network, where:

[0016] The hypernetwork is composed of a stack of basic units that contain all possible image classification operations;

[0017] S2, builds a structure generation network based on a convolutional neural network, samples from a standard Gaussian distribution, obtains the sampled value as the input of the structure generation network, and obtains the output of the structure generation network through forward propagation;

[0018] Then sample the noise from the standard Gaussian distribution;

[0019] The output of the structure generation network and the noise obtained by sampling are summed as the structural parameters of the super network;

[0020] S3, inputting the first part of training images into the hypernetwork to generate predicted category labels;

[0021] Calculating the image classification cross entropy loss based on the predicted category label and the true category label;

[0022] Using a gradient descent method to update the network parameters of the hypernetwork according to the image classification cross entropy loss;

[0023] S4, inputting the second part of training images into the hypernetwork, maximizing the mutual information between the image data and the structural parameters of the hypernetwork, and determining the lower bound of the mutual information, wherein:

[0024] The lower bound of the mutual information is the cross entropy loss of the posterior distribution of the structure parameters and the posterior distribution of the image data. The cross entropy loss is calculated, and the parameters of the structure generation network are updated using the gradient descent method.

[0025] Repeat S2-S4 to iteratively update the network parameters of the super network and the parameters of the structure generation network until convergence, and stack the updated new basic units to construct the target network.

[0026] Preferably, the structure generation network refers to a neural network formed by stacking a convolutional neural network, a rectified linear unit, and a batch normalization layer. The input of the structure generation network is a sample value ∈~N(0,1) obtained from a standard Gaussian distribution;

[0027] The output of the structure generation network is the sample value ∈ input structure generation network The output of the forward propagation is

[0028] The structural parameters of the hypernetwork are obtained by sampling noise ε~N(0,1) from a standard Gaussian distribution and summing it with the output of the structure generation network. That is, the structural parameters of the hypernetwork are composed of the sum of two parts. One part is generated by the structure generation network based on sampling from a standard Gaussian distribution with a mean of 0 and a variance of 1; the other part is the noise sampled from the standard Gaussian distribution.

[0029] Preferably, the mutual information between the image data and the structural parameters of the hypernetwork refers to Where D represents the given training image data, A represents the structural parameters of the hypernetwork, Represents the network parameters describing the data set D given the structural parameters A.

[0030] Preferably, the lower bound of the mutual information is in:

[0031] H(D) represents the information entropy of the training image data D, represents the joint probability distribution of the training data and the structural parameters of the hypernetwork, q θ (D|A) represents the posterior distribution of the approximate real data The variational distribution of .

[0032] Preferably, the cross entropy loss between the posterior distribution of the structural parameters and the posterior distribution of the image data refers to in Represents the conditional probability distribution of the structural parameters of the hypernetwork after given training image data D, q θ (D|A) represents the variational conditional probability distribution of the training image data given the structural parameters of the hypernetwork.

[0033] Preferably, the updated new basic units are stacked to construct a target network, including: obtaining sampling values ​​from a standard Gaussian distribution, inputting a structure generation network, obtaining structural parameters of an output super network, and selecting a specified number of image classification operations with the largest corresponding parameter values ​​from the super network constructed by the basic units according to the structural parameters of the super network, to finally obtain the target network.

[0034] According to a second aspect of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned image classification method for maximizing mutual information when executing the program.

[0035] According to a third aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned image classification method for maximizing mutual information.

[0036] According to a fourth aspect of the present invention, a chip system is provided, comprising a processor, the processor being coupled to a memory, the memory storing program instructions, and the image classification method for maximizing mutual information is implemented when the program instructions stored in the memory are executed by the processor.

[0037] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0038] The above-mentioned image classification method for maximizing mutual information of the present invention is based on given image data and automatically designs and determines the network structure and parameters of the neural network for image classification by maximizing mutual information. It does not require complicated manual design and saves human resources and computing resources.

[0039] The above-mentioned image classification method for maximizing mutual information of the present invention can automatically design an image classification method based on a neural network in a very short time, and can achieve a higher image classification accuracy rate at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0041] Figure 1 A flowchart of an image classification method for maximizing mutual information in one embodiment of the present invention;

[0042] Figure 2 A schematic diagram of a super network of basic cell stacks according to an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of a target structural unit finally obtained by continuously training and iteratively updating the network parameters of the hypernetwork and the parameters of the structure generation network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several variations and improvements may be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0045] The present invention provides an embodiment, an image classification method for maximizing mutual information, comprising:

[0046] Collect training images;

[0047] Maximize the mutual information between the training image and the neural network structure, and automatically determine the network structure and parameters of the neural network;

[0048] The obtained neural network is used to process the image data to be classified to obtain the image classification result.

[0049] As a preferred embodiment, the mutual information between the training image and the neural network structure is maximized, and the network structure and parameters of the neural network are automatically determined, including:

[0050] Constructing a hypernetwork and a structure generation network, performing data processing on them to obtain the network parameters of the hypernetwork and the parameters of the structure generation network, and constructing a target network;

[0051] All training images are input into the target network to generate predicted image category labels. Based on the predicted image category labels and the actual image category labels, the cross entropy loss of image classification is calculated. The target network is trained until convergence for image classification.

[0052] As a preferred embodiment, a hypernetwork and a structure generation network are constructed, data processing is performed on them respectively to obtain network parameters of the hypernetwork and parameters of the structure generation network, and a target network is constructed, including:

[0053] S1, based on all possible operations of image classification, a hypernetwork is constructed using basic units, where the hypernetwork is stacked by basic units that contain all possible image classification operations;

[0054] The hypernetwork is composed of a series of basic units stacked according to certain rules. Each basic unit includes multiple nodes. There are all possible inter-layer connections (edges) between nodes, and all possible image classification operations are defined on the edges, including convolution, pooling, skip-connect, etc. For any two connected nodes, it means that a certain node is used as input, and the output is passed to the connected node through the operation on the connected edge. In the hypernetwork, each basic unit contains two input nodes, four intermediate nodes and one output node. Except that the two input nodes are not connected and the output node is a cascade of four intermediate nodes, all nodes are connected in pairs, forming a total of fourteen possible edges, and each possible edge contains all eight possible operations. The maximum mutual information method obtains the posterior distribution of the structure parameters and the cross entropy loss of the posterior distribution of the image data according to the input image data, iterates the training to update the parameters of the structure generation network, and finally determines the structure of the basic unit (the choice of possible operations) from the hypernetwork.

[0055] like Figure 2 As shown, in one embodiment, the basic unit is Figure 2 As shown, the super network is stacked, in which the basic units are divided into two categories: Normal unit and Reduction unit. The step size of all possible operations in the Normal unit is 1, so that the output is consistent with the input dimension after passing through the Normal unit, and the step size of the operation of the adjacent input node in the Reduction unit is set to 2, so that the input dimension is reduced by half after passing through this unit. This embodiment sets a reasonable overall structure for the super network. After the basic units are obtained through network structure search, they are stacked into the final overall network structure that can be used.

[0056] S2, construct a structure generation network based on a convolutional neural network, sample from a standard Gaussian distribution, obtain the sampled value as the input of the structure generation network, and obtain the output of the structure generation network through forward propagation; then sample from a standard Gaussian distribution to obtain noise; sum the output of the structure generation network and the sampled noise as the structural parameter of the hypernetwork;

[0057] S3, input the first part of training images into the hypernetwork to generate predicted category labels; calculate the image classification cross entropy loss based on the predicted category labels and the true category labels; use the gradient descent method to update the network parameters of the hypernetwork based on the image classification cross entropy loss;

[0058] First, the image data is input into the convolution layer, rectified linear unit, and batch normalization layer of the neural network to obtain a feature map. The convolution kernel slides on the input image to perform a convolution operation, and new image features are extracted through the convolution filter. Multiple convolution kernels (from multiple angles) can be used to extract image features. The obtained feature map is then input into the fully connected layer, and the extracted image features are mapped to the corresponding image category labels to obtain the predicted category label of the image. Based on the true category label of the image, the image classification cross entropy loss is calculated, and then the network parameters of the hypernetwork are updated by the gradient descent method.

[0059] Network parameters refer to the parameters of all image classification operations in all basic units, such as the weight parameters of convolution kernels.

[0060] S4, inputting the second part of the training images into the hypernetwork, maximizing the mutual information between the image data and the structural parameters of the hypernetwork, and determining the lower bound of the mutual information, wherein: the lower bound of the mutual information is the cross entropy loss of the posterior distribution of the structural parameters and the posterior distribution of the image data, calculating the cross entropy loss, and using the gradient descent method to update the parameters of the structure generation network;

[0061] Repeat S2-S4 to iteratively update the network parameters of the super network and the parameters of the structure generation network until convergence, and then stack the updated new basic units to construct the target network.

[0062] As a preferred embodiment, the structure generation network refers to a neural network stacked by a convolutional neural network, a rectified linear unit, and a batch normalization layer. Sampling from the standard Gaussian distribution N(0, 1), the sampling value is ∈~N(0, 1), which is used as the input of the structure generation network. After forward propagation, the output of the structure generation network is Then sample the noise e~N(0,1) from the standard Gaussian distribution and sum it with the output of the structure generation network to obtain Participate in training as the structural parameters of the hypernetwork.

[0063] As a preferred embodiment, the mutual information between the image data and the structural parameters of the hypernetwork is Where D represents the given training image data, A represents the structural parameters of the hypernetwork, Represents the network parameters describing the data set D given the structural parameters A.

[0064] As a preferred embodiment, the lower bound of the mutual information is Where: H(D) represents the information entropy of the training image data D, represents the joint probability distribution of the training data and the structural parameters of the hypernetwork, q θ (D|A) represents the posterior distribution of the approximate real data The variational distribution of .

[0065] As a preferred embodiment, the cross entropy loss of the posterior distribution of the structural parameters and the posterior distribution of the image data refers to in represents the conditional probability distribution of the structural parameters of the hypernetwork after given training image data D, q θ (D|A) represents the variational conditional probability distribution of the training image data given the structural parameters of the hypernetwork.

[0066] As a preferred embodiment, the updated new basic units are stacked to construct a target network, including: obtaining sampling values ​​from a standard Gaussian distribution, inputting a structure generation network, obtaining structural parameters of an output super network, and selecting a specified number of image classification operations with the largest corresponding parameter values ​​from the super network constructed by the basic units according to the structural parameters of the super network, to finally obtain the target network.

[0067] Based on the same concept of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the image classification method for maximizing mutual information is implemented when the processor executes the program.

[0068] Based on the same concept of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that when the program is executed by a processor, the above-mentioned image classification method for maximizing mutual information is implemented.

[0069] Based on the same concept of the present invention, a chip system is provided, including a processor, a coupling between the processor and a memory, and the memory stores program instructions. It is characterized in that when the program instructions stored in the memory are executed by the processor, the image classification method for maximizing mutual information as claimed in the claim is implemented.

[0070] The present invention also provides a flow chart of another preferred embodiment, referring to Figure 1 Image classification methods that maximize mutual information include:

[0071] S1, based on all possible operations of image classification, a hypernetwork is constructed using basic units, where the hypernetwork is stacked by basic units that contain all possible image classification operations;

[0072] S2, constructs a structure generation network based on a convolutional neural network, samples from a standard Gaussian distribution, obtains the sampled value as the input of the structure generation network, obtains the output of the structure generation network through forward propagation, and then samples from a standard Gaussian distribution to obtain noise. The output and the sampled noise are summed as the structural parameters of the hypernetwork;

[0073] S3, input the first part of training images into the hypernetwork to generate predicted category labels; calculate the image classification cross entropy loss based on the predicted category labels and the true category labels; use the gradient descent method to update the network parameters of the hypernetwork based on the image classification cross entropy loss;

[0074] S4, input the second part of training images into the hypernetwork, maximize the mutual information between the image data and the structural parameters of the hypernetwork, derive the lower bound of the mutual information as the cross entropy loss of the posterior distribution of the structural parameters and the posterior distribution of the image data, calculate the cross entropy loss, and use the gradient descent method to update the parameters of the structure generation network;

[0075] In this step, the mutual information between the image data and the structural parameters of the hypernetwork is Where D represents the given training image data, A represents the structural parameters of the hypernetwork, It represents the network parameters that can accurately describe the data set D given the structural parameters A. The formula is equivalent to replace

[0076]

[0077] Where H(D) represents the information entropy of the training image data D, represents the joint probability distribution of the training data and the structural parameters of the hypernetwork, q θ (D|A) represents the posterior distribution of the approximate real data The variational distribution of , since the KL divergence is always greater than or equal to 0, that is Always holds true, and the lower bound of the mutual information is

[0078]

[0079] The cross entropy loss between the posterior distribution of the structural parameters and the posterior distribution of the image data means that for a given training image data, the information entropy H(D) is a constant value that is independent of the model parameters. Therefore, optimizing the lower bound of the mutual information can be further equivalent to optimizing

[0080]

[0081] The formula is the cross entropy loss between the posterior distribution of the structural parameters and the posterior distribution of the image data.

[0082] The cross entropy loss is optimized using the gradient descent method to update the parameters of the structure generation network.

[0083] Repeat the above S2 to S4, continuously iteratively update the network parameters and structural parameters of the super network until the training converges, and finally obtain the target network.

[0084] S5, input all training images into the target network, generate predicted image category labels, calculate the cross entropy loss of image classification according to the predicted image category labels and the real image category labels, train the target network until convergence, and use it for image classification. In this embodiment, the target network is trained based on the criterion of reducing the cross entropy loss, and the training finally obtains the target network with the minimum cross entropy loss.

[0085] like Figure 3 As shown in the figure, after continuous training and iterative updating of network parameters and structural parameters, the target structural unit is finally obtained: the left side is a normal cell (Normal cell), and the right side is a reduction cell (Reduction cell). The target structural unit is output to the image classification module as the result of maximizing mutual information. In order to illustrate the technical effect of the above method, in the embodiment of the present invention, Figure 3 The target structural unit shown in the figure is the converged structure obtained after one epoch of training the image classification method that maximizes mutual information. This method can automatically design the target optimization network for image classification in just 0.007 GPU-days. Most neural networks used for image classification require a long time to train, usually reaching convergence in tens to hundreds of epochs depending on the convergence conditions. The target optimization network achieved an image classification test error (%) of 2.45+0.04 with 3.9M parameters on CIFAR-10. When migrated to CIFAR-100 and ImageNet datasets, it can achieve image classification test error rates of 15.80% and 24.0%, respectively. The automatic design process of this image classification method can be obtained by running on a single Nvidia 1080Ti GPU in about 0.09 hours. Compared with DARTS (9.6 hours) and SNAS (36 hours), this result can significantly improve the design efficiency of image classification methods, while also achieving a high image classification accuracy, ensuring the practicality and stability of image classification methods.

[0086] Specific comparison results can be seen in Tables 1, 2, 3 and 4 below. Table 1 is a comparison of the evaluation of the search results of an embodiment of the present invention (on the CIFAR-10 dataset), Table 2 is a comparison of the evaluation of the search results of an embodiment of the present invention (on the ImageNet dataset), Table 3 is a comparison of the evaluation of the search results of an embodiment of the present invention (on the NAS-Bench-201 benchmark), and Table 4 is a comparison of the evaluation of the search results of an embodiment of the present invention (on the search space of S1-S4). In the table, the lower the error rate, the better the performance, and the smaller the number of parameters and the amount of calculation, the more efficient the structure.

[0087] Table 1

[0088]

[0089] Table 2

[0090]

[0091] Table 3

[0092]

[0093] Table 4

[0094]

[0095] The above experiments show that the image classification method for maximizing mutual information proposed in the embodiments of the present invention can improve the performance of the image classification method under limited computing resources and achieve higher accuracy, and can automatically design and adjust the image classification method based on neural network in a relatively short time, thereby having a wider range of application scenarios.

[0096] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various modifications or variations within the scope of the claims, which does not affect the essence of the present invention. The above preferred features can be used in any combination without conflicting with each other.

Claims

1. An image classification method for maximizing mutual information, characterized in that: include: Collect training images; Maximize the mutual information between the training image and the neural network structure, and automatically determine the network structure and parameters of the neural network; The obtained neural network is used to process the image data to be classified to obtain an image classification result; The collected training images are divided into two parts; the method of maximizing the mutual information between the training images and the neural network structure and automatically determining the network structure and parameters of the neural network includes: Constructing a hypernetwork and a structure generation network, performing data processing on them to obtain the network parameters of the hypernetwork and the parameters of the structure generation network, and constructing a target network; Input all training images into the target network, generate predicted image category labels, calculate the cross entropy loss of image classification based on the predicted image category labels and the actual image category labels, train the target network until convergence, and use it for image classification; The method of constructing a hypernetwork and a structure generation network, performing data processing on them respectively to obtain network parameters of the hypernetwork and parameters of the structure generation network, and constructing a target network includes: S1, constructs basic units based on all possible operations for image classification; The basic unit is then used to construct a super network, where: The hypernetwork is composed of a stack of basic units that contain all possible image classification operations; S2, builds a structure generation network based on a convolutional neural network, samples from a standard Gaussian distribution, obtains the sampled value as the input of the structure generation network, and obtains the output of the structure generation network through forward propagation; Then sample the noise from the standard Gaussian distribution; The output of the structure generation network and the noise obtained by sampling are summed as the structural parameters of the super network; S3, inputting the first part of training images into the hypernetwork to generate predicted category labels; Calculating the image classification cross entropy loss based on the predicted category label and the true category label; Using a gradient descent method, updating the network parameters of the hypernetwork according to the image classification cross entropy loss; S4, inputting the second part of training images into the hypernetwork, maximizing the mutual information between the image data and the structural parameters of the hypernetwork, and determining the lower bound of the mutual information, wherein: The lower bound of the mutual information is the cross entropy loss of the posterior distribution of the structure parameters and the posterior distribution of the image data. The cross entropy loss is calculated, and the parameters of the structure generation network are updated using the gradient descent method. Repeat S2-S4 to iteratively update the network parameters of the super network and the parameters of the structure generation network until convergence, and stack the updated new basic units to construct the target network; The lower bound of the mutual information is in: Represents training image data The information entropy of represents the joint probability distribution of the training data and the structural parameters of the hypernetwork, Represents the posterior distribution used to approximate the true data The variational distribution of .

2. The image classification method for maximizing mutual information according to claim 1, characterized in that: The structure generation network refers to a neural network formed by stacking convolutional neural networks, rectified linear units, and batch normalization layers. ; The input of the structure generation network is a sample value obtained from a standard Gaussian distribution ; The output of the structure generation network is to convert the sampled values Input structure generation network , the output of forward propagation is ; The structural parameters of the hypernetwork are noise sampled from a standard Gaussian distribution. , and summed with the output of the structure generation network to obtain .

3. The image classification method for maximizing mutual information according to claim 1, characterized in that: The mutual information between the image data and the structural parameters of the hypernetwork is ,in Represents the given training image data, represents the structural parameters of the hypernetwork, Represents a given structure parameter In the case of network parameters.

4. The image classification method for maximizing mutual information according to claim 1, characterized in that: The cross entropy loss between the posterior distribution of the structural parameters and the posterior distribution of the image data is ,in Represents the given training image data After that, the conditional probability distribution of the structural parameters of the hypernetwork is Represents the variational conditional probability distribution of the training image data given the structural parameters of the hypernetwork.

5. The image classification method for maximizing mutual information according to claim 1, characterized in that: The updated new basic units are stacked to build a target network, including: The sampling value is obtained from the standard Gaussian distribution and input into the structure generation network to obtain the structural parameters of the output hypernetwork. According to the structural parameters of the hypernetwork, a specified number of image classification operations with the largest corresponding parameter values ​​are selected from the hypernetwork constructed by the basic units to finally obtain the target network.

6. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the image classification method for maximizing mutual information according to any one of claims 1 to 5 is implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the image classification method for maximizing mutual information as described in any one of claims 1 to 5 is implemented.

8. A chip system, comprising a processor, the processor coupled to a memory, the memory storing program instructions, characterized in that: When the program instructions stored in the memory are executed by the processor, the image classification method for maximizing mutual information according to any one of claims 1 to 5 is implemented.

Images