CN114048837A - A Deep Neural Network Model Reinforcement Method Based on Distributed Brain-like Graph - Google Patents

Abstract

The invention discloses a deep neural network model reinforcement method based on a distributed brain-like graph, which generates a backbone brain-like graph by applying a key path to the distributed brain-like graph to improve the propagation process and node behavior on the path. Effective constraints such as constructing a new gradient loss function to weaken the propagation of noise, so as to use graph network metrics such as feature path length and participation coefficient to grow into a new robust brain-like graph structure on the backbone brain-like graph guided by critical paths. Reconstruct the model, improve the robustness of the model, and strengthen the model. The method of the invention uses the brain-like diagram to show the closer connection with the biological neural network, only operates on some key neurons in the neural network, and greatly preserves the integrity of the neural network.

Description

A Deep Neural Network Model Reinforcement Method Based on Distributed Brain-like Graph

technical field

The invention relates to the field of distributed machine learning and artificial intelligence security, in particular to a method for strengthening a deep neural network model based on a distributed brain-like graph.

Background technique

With the substantial improvement of software performance and hardware computing power in modern society, artificial intelligence has been widely used in computer vision, natural language processing, complex network analysis and other fields with good results. However, Christian et al. proposed that adding a misleading perturbation to the original sample image to generate a new sample would cause the model to give a wrong output with high confidence. This newly generated sample is called an adversarial sample, and it poses a potential security threat to deep learning systems such as face recognition systems, automatic verification systems, and autonomous driving systems.

In the past few years, a large number of defense methods have been proposed to improve the robustness of models to adversarial examples to avoid potential dangers in real-world applications. These methods can be roughly classified into adversarial training, input transformation, model architecture transformation, and adversarial sample detection. However, most of these methods above are aimed at the pixel space of the input image, and few of them analyze the impact of adversarial perturbations by studying the inter-layer structure of the model. This is because while it is widely believed that the performance of a neural network depends on its architecture, there is a lack of systematic understanding of the relationship between neural network accuracy and its underlying graph structure. In a recent study, on the one hand, You et al. proposed a new way to represent neural networks as graphs, called relational graphs. The diagram focuses primarily on the exchange of information, not just directed data flow. However, this method can only build the original robust model. Once attacked by adversarial samples, it is difficult to explain and defend from a fine-grained point of view. At the same time, this relationship graph is still not separated from the traditional image network, and lacks connection with biological neural networks. Deep connections in the web. On the other hand, recently Laura et al. simulated the network structure of the human brain by using network topology and modular organization calculation method, and drew a brain network map and counted indicators including characteristic path length and participation coefficient to guide the research, but this kind of The guiding approach lacks an interpretable theoretical basis for the robust performance of neural networks.

At the same time, recently Li et al. proposed a gradient-based influence propagation strategy to obtain critical attack neurons, so as to further construct the critical attack path of the neural network on the computational graph, and through the propagation process and node behavior on the path. Constraints reduce the propagation of noise and improve the robustness of the model. For example, in a social network, false information may bring a huge social threat through the rapid spread among nodes. And nodes with higher information capabilities are more critical than other nodes, are more likely to transmit false information, and are included in the critical path. In order to effectively suppress the spread of false information, an immune strategy is generally adopted, that is, to find and block critical paths in the graph to reduce the spread of false information and improve the security of social networks. But existing graph network representations lack generality and are out of touch with biology and neuroscience.

In view of the above problems, the present invention proposes a method of applying the critical path to the distributed brain-like graph to generate a backbone brain-like graph, and using instructive graph network indicators to grow into a new robust brain-like graph on the backbone brain-like graph A method to reconstruct the graph structure back into the model, thereby improving the robustness of the model.

SUMMARY OF THE INVENTION

In order to further deepen the connection between the deep neural network and the biological neural network, and provide a fine-grained explanation for the representation of the neural network's brain-like graph, the present invention proposes a deep neural network model reinforcement based on the distributed brain-like graph. method, by applying the critical path to the distributed brain-like graph to generate the backbone brain-like graph, so as to use graph network indicators such as characteristic path length and participation coefficient to grow into a new robust brain-like graph structure on the backbone brain-like graph. Reconstruct the model, improve the robustness of the model, and strengthen the model.

In order to achieve the above purpose of the invention, the present invention provides the following technical solutions: a method for strengthening a deep neural network model based on a distributed brain-like graph, comprising the following steps:

(1) Select sample data from the target model dataset;

(2) constructing a target model for the sample data selected in step (1); then training the target model, and finally saving the trained target model;

(3) Define the neural network, and then define the calculation graph of a single neural network, input the target model trained in step (2), and construct the original distributed brain-like graph;

(4) Calculate the influence between each two layers of neurons in the neural network, select key neurons, map the key neurons to the distributed brain-like graph constructed in step (3), and obtain a distributed brain-like graph guided by key paths graph structure;

(5) Define the graph network index, calculate the original distributed brain-like graph obtained in step (3) and the graph network index of the distributed brain-like graph structure guided by the critical path obtained in step (4), respectively, and generate a new class Brain map structure and reconstruct a new target model.

Further, the step (1) is specifically as follows: the target model data set includes n pieces of sample data, which are divided into a type of sample data, and d% of the sample data is extracted from each type of sample data as the training set D of the target model. _train ; where n, a, and d are natural numbers.

Further, the step (2) is specifically:

(2.1) Constructing a target model for the sample data selected in step (1), the target model adopts a distributed structure, and three identical sub-models m ₁ , m ₂ , m ₃ are respectively set for the three RGB features of the image, and Finally, the output model m _out is used to normalize the feature matrix;

(2.2) Set uniform hyperparameters for all the models set in step (2.1) to train the training set D _train set in step (1), specifically: custom setting the number of training epochs, batch size, optimization The optimizer uses stochastic gradient descent, the learning rate is set to the initial cosine learning rate of 0.1, and the loss function Loss _c adds a regularization parameter λ based on the cross entropy function:

表示模型参数，λ表示正则化系数；where p( ) represents the true label of the sample, q( ) represents the predicted probability of the model, x _i represents the input sample,

represents the model parameters, and λ represents the regularization coefficient;

(2.3) Repeat the training until the accuracy rate of the target model converges, and then save the target model obtained by training.

Further, the step (3) is specifically:

为边集合，且每个节点v有一个节点的特征向量W_v；(3.1) Define the neural network: define the graph G=(V, E), where V={v ₁ ,...,v _n } is the node set,

is an edge set, and each node v has a feature vector W _v of a node;

为每前后两层之间有传播关系的两个神经元节点之间的连线，边的权重设为由前一层向后一层传播时对应节点的特征向量矩阵的分量，用公式描述为：(3.2) Defining the computational graph of a single model: Using the forward propagation algorithm, define the graph node set V={v ₁ ,...,v _n } as all neurons and edge sets

It is the connection between two neuron nodes that have a propagation relationship between the two layers before and after each, and the weight of the edge is set as the component of the eigenvector matrix of the corresponding node when it propagates from the previous layer to the next layer, which is described by the formula as :

W _v =[ _wi1 , _wi2 ,..., _wij ]

Among them, for each component w _ij , i represents the subscript of the neuron in the previous layer network connected to the weight, i.e. the location, and j represents the subscript of the neuron in the next layer to which the weight is connected, i.e. the location;

(3.3) Build a distributed brain-like graph: input the target model trained in step (2), first calculate the feature vector of each neuron node of each model, and draw the calculation of each model according to the definition in step (3.3) Finally, the calculation graphs of all drawn sub-models are connected with the calculation graph of the output model with the set weights as the edges to generate an original distributed brain-like graph G _ori .

Further, described step (4) is specifically:

中的单个元素z，用参数

表示第l-1层的第i个神经元

对其的影响值：(4.1) Calculate the influence between neurons in two layers: for the output of the jth neuron F _l ^j of an lth layer

A single element z in , with the parameter

represents the ith neuron of layer l-1

Its impact value:

是第l-1层的第i个神经元的元素的集合，A(·)函数提取在指定位置的元素。The subscript l represents the lth layer, l=1,2,...L, L is the total number of layers of the neural network,

is the set of elements of the ith neuron of the l-1th layer, and the A(·) function extracts the element at the specified position.

表示i,j两个神经元之间的影响值即第l-1层的第i个神经元

对第l层的第j个神经元

的输出

中的每个元素的值之和：use

Represents the influence value between two neurons i, j, i.e. the ith neuron in the l-1th layer

For the jth neuron in the lth layer

Output

The sum of the values of each element in :

(4.2) Select key neurons: Given a sample x, first calculate the loss gradient of the ith neuron of the last convolutional layer L contributing to the model decision:

表示第L层中第i个神经元的输出；in,

represents the output of the ith neuron in the Lth layer;

表示最后一层L中选取到的关键神经元：Then, the loss gradients of each neuron are put together to sort from large to small, and the top k neurons are selected as key neurons, and the

Represents the key neurons selected in the last layer L:

表示每一层取到的关键神经元：The top_k( ) function represents the selection of the first k, and _FL is the set of neurons in the last convolutional layer L. Then, based on the influence of the l-1 layer on the l layer, use

Represents the key neurons taken by each layer:

Finally, use R(x) to represent the key neurons of the sample x in different layers:

(4.3) Constraining the loss gradient: A loss term is obtained by restricting the gradient of key neurons:

Then add the loss term to the cross entropy loss to get the final loss function:

Loss=Loss _c +δLoss _g

where Loss _c is denoted as the cross-entropy loss function in step (2), and δ is the hyperparameter used to balance these loss terms;

(4.4) Mapping the critical path: map the key neurons obtained in step (4.2) to the distributed brain-like graph drawn in step (3.3), and remove the nodes and their edges of non-critical neurons in the brain-like graph to get A distributed brain-like graph structure G _path guided by critical paths.

Further, described step (4) is specifically:

(5.1) Define graph network index: the graph network index includes characteristic path length and participation coefficient; the characteristic path length is specifically the average shortest path length of the network, which is used to measure efficiency; the participation coefficient is a node in the network community A measure of the distribution of connections in ;

(5.2) Growing a new brain-like map structure reconstruction model:

Calculate the graph network indicators defined in the above step (5.1) of the original distributed brain-like graph G _ori obtained in step (3) and the distributed brain-like graph structure G _path guided by the critical path obtained in step (4), and observe The change trend of the index, in which the characteristic path length guides the growth trend of the brain-like map, and the participation coefficient guides the distribution weight ratio of each sub-model, and the generated new brain-like map structure is reconstructed back to the target model to obtain m ₁ ′,m ₂ ′, m ₃ ′ submodel and m _out ′ output model.

The technical idea of the present invention is as follows: the deep neural network model reinforcement method based on the distributed brain-like graph provided by the present invention generates the backbone brain-like graph by applying the key path to the distributed brain-like graph to propagate the process on the path. Make more effective constraints on node behaviors, such as building a new gradient loss function to weaken the propagation of noise, so as to use graph network metrics such as characteristic path length and participation coefficient to grow into a new robust on the backbone brain-like graph guided by critical paths. The brain-like graph structure is reconstructed back to the model to improve the robustness of the model. Finally, three state-of-the-art adversarial attack methods are used to generate adversarial samples and attack the model to verify the improvement of the robust performance of the model.

The beneficial results of the present invention are mainly reflected in: 1) Compared with the traditional critical path method that uses computational graphs and newly emerging relational graphs to select key neurons, the use of brain-like graphs shows closer connections with biological neural networks . 2) Provide a fine-grained explanation for the representation of the neural network's brain-like graph by using the critical path method. And the method of using the critical path to improve the robustness of the model only operates on some key neurons in the neural network, which greatly preserves the integrity of the neural network. 3) Using the distributed structure breaks the network construction method in which each feature of the original image is directly flattened into a matrix with the same weight ratio, and realizes the reference-based weight ratio setting under the guidance of the critical path and graph network indicators.

Description of drawings

In order to illustrate the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative efforts.

Fig. 1 is the flow chart of the inventive method;

Fig. 2 is the schematic diagram of model building brain-like diagram in the present invention;

FIG. 3 is a schematic diagram of selecting key neurons in the relationship diagram in the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and do not limit the protection scope of the present invention.

1 to 3 , the present invention provides a method for strengthening a deep neural network model based on a distributed brain-like graph, including the following steps:

(1) Build the target model dataset, specifically:

The target model data set includes n pieces of sample data, the sample data is divided into a type of sample data, d% of the sample data is extracted from each type of sample data as the training set D _train of the target model, and the remaining samples of each type are The data is used as the test set D _test of the target model; wherein, n, a, and d are natural numbers;

In the embodiment of the present invention, the CIFAR-10 dataset is used to construct the brain-like map and verify its robustness. The CIFAR-10 dataset contains a total of 60,000 RGB color images, each with a size of 32*32, divided into 10 categories, with a total of 6,000 samples in each category. Among them, there are 50,000 training samples and 10,000 test samples. In the present invention, all 50,000 training samples of the CIFAR-10 data set are taken as the training set D _train of the target model, and all 10,000 samples are taken from the test samples as the test set D _test of the target model.

(2) Build and train the target model, including the following sub-steps:

(2.1) Constructing a target model for the sample data selected in step (1), the target model adopts a distributed structure, and three identical sub-models m ₁ , m ₂ , m ₃ are respectively set for the three RGB features of the image, and The output model _mout is finally used to normalize the feature matrix.

In the embodiment of the present invention, three 5-layer MLPs with 512 hidden units are used as the target sub-model structure in the CIFAR-10 dataset, and one 2-layer MLP with 32 hidden units is used as the target output model structure. The input of the MLP is A 3072-dimensional flat vector of CIFAR-10 images (32*32*3), the output is a 10-dimensional prediction, each MLP layer has a ReLU activation function and a BatchNorm regularization layer.

(2.2) Set unified hyperparameters for all target models set in step (2.1), and train the training set D _train set in step (1): customize training epoch times, batch size, optimizer, learning rate And the loss function, using stochastic gradient descent, the learning rate is set to the cosine learning rate of the initial 0.1, and the loss function Loss _c adds the regularization parameter λ on the basis of the cross entropy function:

表示模型参数，λ表示正则化系数。where p( ) represents the true label of the sample, q( ) represents the predicted probability of the model, x _i represents the input sample,

represents the model parameters, and λ represents the regularization coefficient.

In the embodiment of the present invention, uniform hyperparameters are set: the number of training epochs is 200, the batch size is 128, stochastic gradient descent (SGD) is used, the learning rate is set to the initial cosine learning rate of 0.1, and the loss function is based on the cross-entropy function. The regularization parameter is added to λ, and the above loss function Loss _c calculation formula is obtained.

(2.3) Repeat the training until the accuracy rate of the target model basically converges and no longer improves; then save the target model obtained by training.

(3) Define the neural network, then define the calculation graph of a single neural network, input the target model trained in step (2), and construct a distributed brain-like graph, including the following sub-steps:

为边集合，且每个节点v有一个节点的特征向量W_v。(3.1) Define the neural network: define the graph G=(V, E), where V={v ₁ ,...,v _n } is the node set,

is a set of edges, and each node v has a eigenvector W _v of the node.

(3.2) Define the computation graph of a single model:

为每前后两层之间有传播关系的两个神经元节点之间的连线，边的权重设为由前一层向后一层传播时对应节点的特征向量矩阵的分量，以全连接网络为例，用公式描述为：Using forward propagation algorithm, define graph node set V={x ₁ ,...,v _n } as all neurons, edge set

It is the connection between the two neuron nodes that have a propagation relationship between the two layers before and after each, and the weight of the edge is set as the component of the eigenvector matrix of the corresponding node when it propagates from the previous layer to the next layer, with a fully connected network. For example, the formula is described as:

W _v =[ _wi1 , _wi2 ,..., _wij ]

For each component w _ij , i represents the subscript of the neuron in the previous layer network to which the weight is connected, that is, the location, and j represents the subscript of the neuron in the next layer to which the weight is connected, i.e. the location. Generally, in a fully connected network, each neuron of the previous layer of network has connections with all the neurons of the next layer, that is, from the 1st to the jth.

(3.3) Build a distributed brain-like graph:

Load the target model trained in step (2), first calculate the feature vector of each neuron node of each model, draw the calculation graph of each model according to the definition of step (3.2), and finally put all drawn sub The computational graph of the model is connected with the computational graph of the output model with the set weights as the edges to generate a primitive distributed brain-like graph G _ori .

(4) Constraining the critical path, which specifically includes the following sub-steps:

(4.1) Calculate the influence between two layers of neurons:

For any model, in order to construct the critical attack path, it is necessary to extract the critical attack neurons in each layer and connect them. The first step is to calculate the influence of the neurons in the previous layer on the neurons in the next layer by back-propagation.

的输出

中的单个元素z，用参数

表示第l-1层的第i个神经元

对其的影响值：Specifically, the influence of one neuron is the sum of the absolute values of the gradients of the elements at each position with respect to the other neuron. For the jth neuron of an lth layer

Output

A single element z in , with the parameter

represents the ith neuron of layer l-1

Its impact value:

是第l-1层的第i个神经元的元素的集合，A(·)函数提取在指定位置的元素。The subscript l represents the lth layer, l=1,2,...L, L is the total number of layers of the neural network,

is the set of elements of the ith neuron of the l-1th layer, and the A(·) function extracts the element at the specified position.

表示i,j两个神经元之间的影响值即第l-1层的第i个神经元

对第l层的第j个神经元

的输出

中的每个元素的值之和：use

Represents the influence value between two neurons i, j, i.e. the ith neuron in the l-1th layer

For the jth neuron in the lth layer

Output

The sum of the values of each element in :

(4.2) Select key neurons:

Given a sample x, first derive the gradient of the loss that the ith neuron of the last convolutional layer L contributes to the model decision:

表示第L层中第i个神经元的输出。in

represents the output of the ith neuron in the Lth layer.

表示最后一层L中选取到的关键神经元：Then, the loss gradients of each neuron are put together to sort from large to small, and the top k neurons are selected as key neurons, and the

Represents the key neurons selected in the last layer L:

表示每一层取到的关键神经元：The top_k(·) function means to select the first k, and _FL is the set of neurons of the last convolutional layer L, and then based on the influence of the l-1 layer on the l layer, the formula of step 4.1) is used to recursively obtain the first few layers. key neurons, with

Represents the key neurons taken by each layer:

Finally, use R(x) to represent the key neurons of the sample x in different layers:

(4.3) Constrain the loss gradient:

An intuitive way to constrain the critical path in the face of adversarial attacks is to limit the gradient of the loss to reduce the influence of these neurons. A loss term can be directly obtained by limiting the gradient of key neurons:

Then add the loss term to the cross entropy loss to get the final loss function:

Loss=Loss _c +δLoss _g

where Loss _c represents the cross-entropy loss function in step (2), and δ is the hyperparameter used to balance these loss terms.

(4.4) Mapping the critical path:

Map the key neurons obtained in step (4.2) to the distributed brain-like graph drawn in step (3.3), and remove the nodes and edges of non-critical neurons in the brain-like graph to obtain a distributed brain-like graph guided by critical paths. Brain-like graph structure G _path .

(5) Reconstructing the model under the guidance of graph network indicators, including the following sub-steps:

(5.1) Define various graph network metrics:

①Characteristic path length: The feature path length is an index to measure the efficiency, which is defined as the average shortest path length of the network. The distance matrix used to calculate the shortest path must be a link length matrix, usually obtained by mapping from weights to lengths. Here, the most commonly used weighted path length is used as the calculation standard, and the formula is as follows:

WPL=∑w _ij *l

where w _ij is the edge weight defined in step (3.2), l is the subscript in step (4.1), representing the lth layer, l=1,2,...L, L is the total number of layers of the neural network, here generally Set l to the layer where the neuron with index i in w _ij is located.

②Participation coefficient: The participation coefficient is a measure of the connection distribution of a node in the network community. When the participation coefficient is 0, the connection of a node is completely limited to its block. The closer the participation coefficient is to 1, the more uniform the distribution of nodes' connections between blocks. Mathematically, the participation coefficient P of node i is:

where S _is the sum of the connection weights from node i to nodes in block c, S _i is the strength of node i, and C is the total number of blocks. Here each sub-model is set as a block;

(5.2) Growing a new brain-like map structure reconstruction model

Calculate the two indicators in the above (5.1) of the original distributed brain-like graph G _ori obtained in step (3) and the distributed brain-like graph structure G _path guided by the critical path obtained in step (4). The changing trend, in which the characteristic path length guides the growth trend of the brain-like map, and the participation coefficient guides the distribution weight ratio of each sub-model, and finally generates a new brain-like map structure after growth, and converts the generated new brain-like map structure back to model, that is, reconstructed back to the target model to obtain robust target models of m ₁ ′, m ₂ ′, m ₃ ′ and m _out ′.

(6) Conduct adversarial attacks on the target model to generate adversarial samples, and use the attack model to verify the improvement of the robustness of the model:

The embodiment of the present invention adopts a variety of adversarial attack methods, including FGSM attack, CW attack and PGD attack. Each attack randomly selects 1000 generated adversarial samples in each dataset to attack. The three attacks set different parameters. For FGSM attack, set parameter ε=2; for CW attack, use L ₂ norm attack, set initial value c=0.01, confidence degree k=0, iteration number epoch=200; For the PGD attack, set the parameter ε=2, the step size α=ε/10, and the number of iterations epoch=20.

Model robustness evaluation index; when subjected to adversarial attacks, the model often uses accuracy as an evaluation index for robust performance.

Accuracy: Accuracy represents the ratio of the number of samples correctly classified by the classifier to the total number of samples for a given test data set

Among them, TP means that the positive class is judged as the positive class, FP means that the negative class is judged as the positive class, FN means that the positive class is judged as the negative class, TN means that the negative class is judged as the negative class, the lower the accuracy rate, the more robust performance the better. The accuracy of the robust target model of the CIFAR-10 dataset under three attacks is improved by an average of 42.3% compared with the original target model.

The above-mentioned specific embodiments describe in detail the technical solutions and beneficial effects of the present invention. It should be understood that the above-mentioned embodiments are only the most preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, additions and equivalent substitutions made within the scope shall be included within the protection scope of the present invention.

Claims (6)

1. a deep neural network model reinforcement method based on distributed brain-like graph, is characterized in that, comprises the following steps: (1) Select sample data from the target model dataset; (2) constructing a target model for the sample data selected in step (1); then training the target model, and finally saving the trained target model; (3) Define the neural network, and then define the calculation graph of a single neural network, input the target model trained in step (2), and construct the original distributed brain-like graph; (4) Calculate the influence between each two layers of neurons in the neural network, select key neurons, map the key neurons to the distributed brain-like graph constructed in step (3), and obtain a distributed brain-like graph guided by key paths graph structure; (5) Define the graph network index, calculate the original distributed brain-like graph obtained in step (3) and the graph network index of the distributed brain-like graph structure guided by the critical path obtained in step (4), respectively, and generate a new class Brain map structure and reconstruct a new target model. 2. The deep neural network model reinforcement method based on distributed brain-like map according to claim 1, wherein the step (1) is specifically: the target model data set comprises n pieces of sample data, which are divided into For the sample data of type a, d% of the sample data is extracted from each type of sample data as the training set D _train of the target model; wherein, n, a, and d are natural numbers. 3. the deep neural network model reinforcement method based on distributed brain-like graph according to claim 1, is characterized in that, described step (2) is specifically: (2.1) Construct a target model for the sample data selected in step (1). The target model adopts a distributed structure, and three identical sub-models m ₁ , m ₂ , m ₃ are respectively set for the three RGB features of the image, and Finally, the output model m _out is used to normalize the feature matrix; (2.2) Set uniform hyperparameters for all the models set in step (2.1) to train the training set D _train set in step (1), specifically: custom setting the number of training epochs, batch size, optimization The optimizer uses stochastic gradient descent, the learning rate is set to the initial cosine learning rate of 0.1, and the loss function Loss _c adds a regularization parameter λ based on the cross entropy function:

where p( ) represents the true label of the sample, q( ) represents the predicted probability of the model, x _i represents the input sample,

represents the model parameters, and λ represents the regularization coefficient; (2.3) Repeat the training until the accuracy rate of the target model converges, and then save the target model obtained by training. 4. the deep neural network model reinforcement method based on distributed brain-like graph according to claim 1, is characterized in that, described step (3) is specifically: (3.1) Define the neural network: define the graph G=(V, E), where V={v ₁ , . . . , v _n } is the set of nodes,

is an edge set, and each node v has a feature vector W _v of a node; (3.2) Defining the computational graph of a single model: Using the forward propagation algorithm, define the graph node set V={v ₁ ,..., v _n } as all neurons and edge sets

It is the connection between two neuron nodes that have a propagation relationship between the two layers before and after each, and the weight of the edge is set as the component of the eigenvector matrix of the corresponding node when it propagates from the previous layer to the next layer, which is described by the formula as : W _v =[ _wi1 , _wi2 ,..., _wij ] Among them, for each component w _ij , i represents the subscript of the neuron in the previous layer network connected to the weight, i.e. the location, and j represents the subscript of the neuron in the next layer to which the weight is connected, i.e. the location; (3.3) Build a distributed brain-like graph: input the target model trained in step (2), first calculate the feature vector of each neuron node of each model, and draw the calculation of each model according to the definition in step (3.3) Finally, the calculation graphs of all drawn sub-models are connected with the calculation graph of the output model with the set weights as the edges to generate an original distributed brain-like graph G _ori . 5. the deep neural network model reinforcement method based on distributed brain-like graph according to claim 1, is characterized in that, described step (4) is specifically: (4.1) Calculate the influence between neurons in two layers: for the output of the jth neuron F _l ^j of an lth layer

A single element z in , with the parameter

represents the ith neuron of layer l-1

Its impact value:

The subscript l represents the lth layer, l=1, 2,...L, L is the total number of layers of the neural network,

is the set of elements of the ith neuron of the l-1th layer, and the A(·) function extracts the element at the specified position. use

Represents the influence value between two neurons i, j, i.e. the ith neuron in the l-1th layer

For the jth neuron in the lth layer

Output

The sum of the values of each element in :

(4.2) Select key neurons: Given a sample x, first calculate the loss gradient of the ith neuron of the last convolutional layer L contributing to the model decision:

in,

represents the output of the ith neuron in the Lth layer; Then, the loss gradients of each neuron are put together for sorting from large to small, and the top k neurons are selected as key neurons, and the

Represents the key neurons selected in the last layer L:

The top_k( ) function represents the selection of the first k, and _FL is the set of neurons in the last convolutional layer L. Then, based on the influence of the l-1 layer on the l layer, use

Represents the key neurons taken by each layer:

Finally, use R(x) to represent the key neurons of the sample x in different layers:

(4.3) Constraining the loss gradient: A loss term is obtained by restricting the gradient of key neurons:

Then add the loss term to the cross entropy loss to get the final loss function: Loss=Loss _c +δLoss _g where Loss _c is denoted as the cross-entropy loss function in step (2), and δ is the hyperparameter used to balance these loss terms; (4.4) Mapping the critical path: map the key neurons obtained in step (4.2) to the distributed brain-like graph drawn in step (3.3), and remove the nodes and edges of non-critical neurons in the brain-like graph to get A distributed brain-like graph structure G _path guided by critical paths. 6. The deep neural network model reinforcement method based on distributed brain-like graph according to claim 1, is characterized in that, described step (4) is specifically: (5.1) Define graph network index: the graph network index includes characteristic path length and participation coefficient; the characteristic path length is specifically the average shortest path length of the network, which is used to measure efficiency; the participation coefficient is a node in the network community A measure of the distribution of connections in ; (5.2) Growing a new brain-like map structure reconstruction model: Calculate the original distributed brain-like graph G _ori obtained in step (3) and the distributed brain-like graph structure G _path with critical path guidance obtained in step (4) respectively. The change trend of the index, in which the characteristic path length guides the growth trend of the brain-like map, and the participation coefficient guides the distribution weight ratio of each sub-model. The generated new brain-like map structure is reconstructed back to the target model to obtain m ₁ ′, m ₂ ', _m3 ' submodel and _mout ' output model.

Images