A method, medium and device for measuring the redundancy reliability of a deep neural network

Abstract

The present invention provides a method, medium and device for measuring the redundant reliability of a deep neural network. The layer nodes of the deep neural network are clustered based on weight information, the geometric center theory is used to identify the geometric center of each layer of the deep neural network, a network reliability calculation model based on a fault injection strategy is constructed, and the redundant reliability of the network is extracted based on an adaptive convergence strategy; a medium executes a deep neural network redundant reliability measurement program for measuring the redundant reliability of a deep neural network based on the geometric center theory; a computer device stores a computer program for implementing the method for measuring the redundant reliability of a deep neural network based on the geometric center theory. The present invention helps to improve the recognition accuracy and speed up the calculation speed; while realizing the calculation acceleration, it effectively quantifies the influence of different outputs on the network reliability, helps relevant personnel to timely understand and master the corresponding relationship between the network redundancy and reliability, and provides an effective reference and basis for further decision-making.

Description

Technical Field

[0001] The present invention relates to the technical field of computer systems based on a specific computational model, and particularly relates to a method, medium, and device for measuring the redundant reliability of a deep neural network based on the geometric center theory in the field of reliability evaluation of an intelligent computing system. Background Art

[0002] In recent years, deep neural networks have been widely applied in all aspects of our lives, such as handwritten digit recognition and driverless driving.

[0003] However, due to the lack of support from credible theories, in the process of network construction, it mainly relies on experience to determine its scale, such as network depth and layer width, resulting in excessive redundant design to meet the requirements of reliability. Structural redundancy does help to improve the reliability level of the network, but it also easily causes an increase in training time and an increase in network power consumption. Therefore, it is necessary to construct a moderate redundant structure to balance the reliability and training overhead of the deep neural network.

[0004] Currently, the methods for analyzing the redundant reliability of deep neural networks mainly include evaluation methods based on fault injection, measurement methods for fault-tolerant computing, and calculation methods based on probability models, etc. The evaluation method based on fault injection quantifies the reliability level of the network by randomly injecting faults into the network and using statistical methods. It can often achieve high computational accuracy, but it is prone to generating huge time overhead. The measurement method for fault-tolerant computing evaluates the reliability of the network by continuously increasing the number of faulty nodes in a specified area to test the network's response to faults. However, since the analysis is only carried out in the specified area each time, this method fails to effectively measure the reliability level of the entire network. The method based on probability models quantifies the uncertain behavior characteristics of nodes through stochastic theory and quantifies the reliability level of the network by using a competitive model. It can usually achieve the purpose of fast calculation, but due to the failure to effectively measure the impact of fault propagation on the output results, it is prone to accuracy loss.

[0005] In summary, the above methods do not clearly show the corresponding relationship between the redundancy and reliability of the network, which is not conducive to optimizing the network reliability and training overhead, nor is it conducive to promoting the realization of the "carbon neutrality" goal. Summary of the Invention

[0006] To overcome the above-mentioned deficiencies of existing methods, the present invention provides an optimized method, medium, and device for measuring the redundancy and reliability of deep neural networks. The geometric center theory is used to measure the correspondence between the redundancy and reliability of deep neural networks, facilitating designers to construct networks that meet application scenarios or helping users select networks that meet requirements. It can not only evaluate the reliability level of the model under different redundancies but also be used to guide designers to implement network compression or pruning.

[0007] The technical solution adopted by the present invention is a method for measuring the redundancy and reliability of a deep neural network based on the geometric center theory. The method clusters the layer nodes of the deep neural network based on weight information, uses the geometric center theory to identify the geometric center of each layer of the deep neural network, constructs a network reliability calculation model based on a fault injection strategy, and extracts the redundancy and reliability of the network based on an adaptive convergence strategy.

[0008] Preferably, the method includes the following steps:

[0009] Step 1: Read the deep neural network model M and extract associated parameters;

[0010] Step 2: Cluster the weights of each layer of nodes in the deep neural network M;

[0011] Step 3: Mark the redundant nodes of each layer in the deep neural network M;

[0012] Step 4: Construct a network reliability calculation model based on a fault injection strategy and calculate the redundancy and reliability of the deep neural network M.

[0013] Preferably, the associated parameters include the number of layers L of the deep neural network M, the weight set W of the layer nodes, the node failure probability P, the reliability R of M under P M , the sample data set V, the maximum number of times E for random sample extraction max , and the redundancy S of the model;

[0014] where W = {W1, W2,.., W L}, and taking W l as the node weight vector of the l-th layer, where l is a positive integer from 1 to L.

[0015] Preferably, the step 2 includes the following steps:

[0016] Step 2.1: Based on the probability distribution of the layer node weights, use a Gaussian kernel function to obtain the number of clusters k of the nodes W l in the l-th layer;

[0017] Step 2.1: Use the K-means algorithm to cluster the W l of the nodes in the l-th layer into k clusters and denote them as Wl,1 ,W l,2 ,..,W l,k 。

[0018] Preferably, step 3 includes the following steps:

[0019] Step 3.1: Initialize the loop variable j = 1;

[0020] Step 3.2: Identify the geometric center of W l,j and denote it as W GM ,

[0021]

[0022] The function argmin returns the element with the minimum sum of Euclidean distances from the n elements in W l,j , where is the i-th element in W l,j , d is the dimension of, R d is a real number in d dimensions, n is the number of elements in W l,j , and ||*||2 refers to the 2-norm;

[0023] Step 3.3: Obtain the element with the smallest Euclidean distance from W GM found first using Equation (2)

[0024]

[0025] Identify the redundant network node corresponding to W l,j with this;

[0026] Step 3.4: According to the preset redundancy S, use Equation (2) to mark the S × len(W l,j ) weights corresponding to the redundant nodes in and place them in the set W l,j , taking W l,j-redundance as the weight set corresponding to the redundant nodes in the j-th classification of the l-th layer; l,j-redundance

[0027] Step 3.5: If j < k, then j = j + 1 and return to step 3.2; otherwise, proceed to the next step;

[0028] Step 3.6: Set the values of all elements in the weight set W l-redundance corresponding to the redundant nodes in the l-th layer to 0, W l-redundance = {0};

[0029] Step 3.7: Mark the weights corresponding to the non-redundant nodes in the l-th layer using Equation (3) and place them in the set W l-key ;​

[0030] W l-key = W l -W l-redundance (3)

[0031] Step 3.8: Denote the updated model of M as M updated .

[0032] Preferably, the said step 4 includes the following steps:

[0033] Step 4.1: Initialize the loop variable E = 0;

[0034] Step 4.2: Randomly extract a sample x from the sample data set V, extract the attribution probability vector vector1 of x based on M from the output end of softmax, and give the ranking rank_golden of the probability of the category of x through vector1; where softmax is the normalized exponential function;

[0035] Step 4.3: Perform fault injection on the set W of the model M according to the given node fault probability P updated and extract the attribution probability vector vector2 of x based on M from the output end of softmax, and then give the ranking rank_fault of the probability of the category of x through vector2; l-key up4ated up4ated up4ated

[0036] Step 4.4: Let rank_golden1 and rank_fault1 correspond to the first elements of rank_golden and rank_fault respectively. If rank_golden1 = rank_fault1, then proceed to the next step; otherwise, return to step 4.2;

[0037] Step 4.5: Calculate the output reliability R1 and cosine similarity R2 of M under the input x according to equations (4) and (5), updated

[0038]

[0039]

[0040] where K is the number of classifications of the output layer of M updated and respectively represent the i-th elements extracted from rank_golden and rank_fault, vector1 i and bector2 i ​​represent the i-th elements extracted from vector1 and vector2 respectively, and ||*||1 is the 1-norm;

[0041] Step 4.6: E = E + 1, calculate the reliability R of M under the input x according to Equation (6) updated of M x ,

[0042] R x = αR1 + βR2 (6)

[0043] where α and β are the importance coefficients of R1 and R2 respectively, and α ∈ [0, 1], β ∈ [0, 1];

[0044] Step 4.7: If E > 1, obtain the unbiased estimate of M using Equation (7) updated and proceed to the next step, otherwise, return to Step 4.2;

[0045]

[0046] where R ′ represents the average value of the reliability of M under randomly extracting E samples from V updated ;

[0047] Step 4.8: Obtain the redundant reliability R of M according to Equation (8);

[0048] R = R M - R′ (8)

[0049] Step 4.9: If R satisfies convergence or E ≥ E max , stop the iteration and proceed to the next step, otherwise return to Step 4.2;

[0050] Step 4.10: Output the redundant reliability R of M under the redundancy S.

[0051] Preferably, in Step 4.9, Equation (9) is used to satisfy convergence,

[0052]

[0053] where ε is the agreed convergence parameter and ε ∈ [0.01, 0.05].

[0054] A computer-readable storage medium, on which a program for measuring the redundant reliability of a deep neural network is stored. When the program for measuring the redundant reliability of the deep neural network is executed by a processor, the method for measuring the redundant reliability of the deep neural network is implemented.

[0055] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method for measuring the redundancy reliability of the deep neural network is implemented.

[0056] The present invention provides an optimized method, medium, and device for measuring the redundancy reliability of a deep neural network. The method clusters the layer nodes of the deep neural network based on weight information, uses the geometric center theory to identify the geometric center of each layer of the deep neural network, constructs a network reliability calculation model based on a fault injection strategy, and extracts the redundancy reliability of the network based on an adaptive convergence strategy; the medium executes a deep neural network redundancy reliability measurement program for the method of measuring the redundancy reliability of the deep neural network based on the geometric center theory; the computer device stores a computer program for implementing the method of measuring the redundancy reliability of the deep neural network based on the geometric center theory.

[0057] The beneficial effect of the present invention is that the identification of network redundant nodes based on the geometric center theory and combined with the clustering algorithm helps to improve the identification accuracy and speed up the calculation speed; the application of the adaptive hierarchical reliability calculation strategy effectively quantifies the impact of different outputs on the network reliability while achieving calculation acceleration, which helps relevant personnel to timely understand and master the corresponding relationship between the redundancy and reliability of the network, providing effective reference and basis for further decisions such as pruning and network structure selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] The following further describes the present invention in detail with reference to embodiments, but the protection scope of the present invention is not limited thereto.

[0060] The present invention relates to a method for measuring the redundancy reliability of a deep neural network based on the geometric center theory. The method clusters the layer nodes of the deep neural network based on weight information, uses the geometric center theory to identify the geometric center of each layer of the deep neural network, constructs a network reliability calculation model based on a fault injection strategy, and extracts the redundancy reliability of the network based on an adaptive convergence strategy.

[0061] In the present invention, the geometric center theory combined with the clustering algorithm is the main technical means for network redundancy reliability analysis. Through the identification of network redundant nodes, the fault injection of non-redundant nodes, and the reliability evaluation based on statistics, the corresponding relationship between the network redundancy and reliability is obtained. This method helps to identify the redundant nodes of the network and provides a reference basis for carrying out lightweight strategies.

[0062] Embodiment 1

[0063] The method includes the following steps:

[0064] Step 1: Read the deep neural network model M and extract the associated parameters;

[0065] The associated parameters include the number of layers L of the deep neural network M, the weight set W of the layer nodes, the node failure probability P, and the reliability R of M under P M , the sample data set V, and the maximum number of times E of random sampling of samples max , and the redundancy S of the model;

[0066] where W = {W1, W2,.., W L}, and taking W l as the node weight vector of the l-th layer, where l is a positive integer from 1 to L.

[0067] In this embodiment, the number of layers L of the deep neural network M and the weight set W of the layer nodes are closely related to the preprocessing of the method, and are used to cluster based on the weights of the nodes in each layer of the deep neural network M and obtain redundant nodes based on the clustering results; with the node failure probability P and the reliability R of M under P M perform failure input on non-redundant nodes and judge the redundancy reliability of the deep neural network M at this time based on the obtained reliability value.

[0068] Step 2: Cluster the weights of the nodes in each layer of the deep neural network M;

[0069] In this embodiment, the clustering of the weights of the nodes in each layer of the deep neural network M is easy for those skilled in the art to understand. Common clustering methods include but are not limited to K-Means clustering, mean shift clustering, DBSCAN, maximum expectation clustering of GMM, etc.

[0070] Step 3: Mark the redundant nodes in each layer of the deep neural network M;

[0071] In this embodiment, Step 3 includes the following steps:

[0072] Step 3.1: Initialize the loop variable j = 1; use the loop variable to implement the loop marking of the redundant nodes in each of the k clusters of the l-th layer;

[0073] Step 3.2: Identify the geometric center of W l,j with Equation (1) and denote it as W GM ,

[0074]

[0075] The function argmin returns the element with the smallest sum of Euclidean distances of the n elements in W l,j as the geometric median, where, For the i-th element in W l,j and d is the dimension of d , R is a d-dimensional real number, n is the number of elements in W l,j , and ||*||2 denotes the 2-norm;

[0076] Step 3.3: Obtain the element with the smallest Euclidean distance to W first according to Equation (2) GM

[0077]

[0078] Use this to identify the redundant network node corresponding to W l,j ;

[0079] Step 3.2 realizes the marking of the geometric center in W l,j and obtains the element with the smallest Euclidean distance based on this as the redundant network node.

[0080] Step 3.4: According to the preset redundancy S, use Equation (2) to mark S×len(W l,j ) weights corresponding to the redundant nodes in W l,j and place them into the set W l,j-redundance , where W l,j-redundance is the set of weights corresponding to the redundant nodes in the j-th classification of the l-th layer;

[0081] Step 3.5: If j < k, then j = j + 1 and return to Step 3.2, otherwise, proceed to the next step; traverse all clusters in the current layer until all clusters are processed;

[0082] Step 3.6: Set the values of all elements in the set W l-redundance of the weights corresponding to the redundant nodes in the l-th layer to 0, W l-redundance = {0};

[0083] Step 3.7: Mark the weights corresponding to the non-redundant nodes in the l-th layer according to Equation (3) and place them into the set W l-key ;

[0084] W l-key = W l - W l-redundance (3)

[0085] Step 3.8: Denote the updated model of M as M updated , where the update here means obtaining the model with the weights set to 0, that is, the model after removing the node weight redundancy S, namely, the benchmark model.

[0086] ​The subsequent step 4 processing is for non-redundant nodes.

[0087] Step 4: Construct a network reliability calculation model based on the fault injection strategy, and calculate the redundant reliability of the deep neural network M.

[0088] The said step 4 includes the following steps:

[0089] Step 4.1: Initialize the loop variable E = 0; the loop variable is preset with the maximum number of times E for random sample extraction max as a limit;

[0090] Step 4.2: Randomly extract a sample x from the sample data set V, and extract the attribution probability vector vector1 of x based on M from the output end of softmax, and give the ranking rank_golden of the probability of the category of x through vector1; where softmax is the normalized exponential function;

[0091] Step 4.3: Implement fault injection on the set W updated of the model M l-key according to the given node fault probability P, and extract the attribution probability vector vector2 of x based on M updated from the output end of softmax, and then give the ranking rank_fault of the probability of the category of x through vector2;

[0092] Step 4.4: Let rank_golden1 and rank_fault1 correspond to the first elements of rank_golden and rank_fault respectively. If rank_golden1 = rank_fault1, proceed to the next step; otherwise, return to step 4.2; here, rank_golden1 is the final classification category of the network, so it is only judged once. If they are the same, proceed to the next step; otherwise, return to step 4.2 to re-extract the sample.

[0093] Step 4.5: Calculate the output reliability R1 and cosine similarity R2 of M updated under the input x with equations (4) and (5),

[0094]

[0095]

[0096] where K is the number of classifications of the output layer of M updated and and respectively represent the i-th elements extracted from rank_golden and rank_fault, vector1 i and vector2i represent the i-th elements extracted from vector1 and vector2 respectively, ||*||1 is the 1-norm;

[0097] Step 4.6: E = E + 1, calculate the reliability R of M under the input x according to Equation (6) up4ated of M x ,

[0098] R x = αR1 + βR2 (6)

[0099] where α and β are the importance coefficients of R1 and R2 respectively, and α ∈ [0, 1], β ∈ [0, 1];

[0100] Step 4.7: If E > 1, obtain the unbiased estimate of M using Equation (7) updated and proceed to the next step, otherwise, return to Step 4.2;

[0101]

[0102] where R ′ represents the average value of the reliability of M under randomly sampling E times from V updated ;

[0103] Step 4.8: Obtain the redundant reliability R of M according to Equation (8);

[0104] R = R M - R′ (8)

[0105] Step 4.9: If R satisfies convergence or E ≥ E max , stop the iteration and proceed to the next step, otherwise return to Step 4.2;

[0106] Step 4.10: Output the redundant reliability R of M under the redundancy S.

[0107] Embodiment 2

[0108] Based on Embodiment 1, Step 2 includes the following steps:

[0109] Step 2.1: Based on the probability distribution of the layer node weights, use the Gaussian kernel function to obtain the number of clusters k of the layer l node W l ;

[0110] Step 2.1: Use the K-means algorithm to cluster the layer l node W l into k clusters, and denote them as W l,1 , W l,2 ,.., W l,k .

[0111] In this embodiment, the K-means algorithm is used to obtain the clustering result.

[0112] Embodiment 3

[0113] Based on Embodiment 1, in step 4.9, Equation (9) is considered to satisfy convergence.

[0114]

[0115] where ε is a predefined convergence parameter and ε ∈ [0.01, 0.05].

[0116] A computer-readable storage medium stores a deep neural network redundancy reliability measurement program. When the deep neural network redundancy reliability measurement program is executed by a processor, the deep neural network redundancy reliability measurement method based on the geometric center theory is implemented.

[0117] To implement the above embodiment, a computer-readable storage medium is proposed. The computer-readable storage medium stores a deep neural network redundancy reliability measurement program. When the deep neural network redundancy reliability measurement program is executed by a processor, the deep neural network redundancy reliability measurement method based on the geometric center theory is implemented.

[0118] According to the computer-readable storage medium of the embodiment of the present invention, through the deep neural network redundancy reliability measurement program, when the processor executes the deep neural network redundancy reliability measurement program, the deep neural network redundancy reliability measurement method based on the geometric center theory as described above is implemented, thereby solving the problem of the lack of correspondence between the redundancy and reliability of the network, which is not conducive to optimizing the network reliability and training cost. It can evaluate the reliability level of the model under different redundancy degrees and can also be used to guide designers to implement network compression or pruning.

[0119] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the deep neural network redundancy reliability measurement method based on the geometric center theory is implemented.

[0120] To implement the above embodiment, a computer device is proposed, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the deep neural network redundancy reliability measurement method based on the geometric center theory as described above is implemented. Similarly, it is used to solve the problem of the lack of correspondence between the redundancy and reliability of the network, which is not conducive to optimizing the network reliability and training cost. It can evaluate the reliability level of the model under different redundancy degrees and can also be used to guide designers to implement network compression or pruning.

[0121] Those skilled in the art will understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) that contain computer-usable program code.

[0122] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or combinations of blocks.

[0123] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or combinations of blocks.

[0124] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are performed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or combinations of blocks.

[0125] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concepts. Therefore, the appended claims are intended to be construed to include the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0126] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and equivalent technologies, the present invention also intends to include these modifications and variations.

[0127] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept of the present invention.

Claims

1. A method for measuring the redundant reliability of a deep neural network, characterized in that: The method clusters the layer nodes of a deep neural network based on weight information, uses the geometric center theory to identify the geometric center of each layer of the deep neural network, constructs a network reliability calculation model based on a fault injection strategy, and extracts the redundant reliability of the network based on an adaptive convergence strategy. The method includes the following steps: Step 1: Read the deep neural network model M and extract the associated parameters; the associated parameters include the number of layers L of the deep neural network M, the weight set W of the layer nodes, the node failure probability P, and the reliability R of M under P M , the sample data set V, and the maximum number of times E for random sampling of samples max , and the redundancy S of the model; where W = {W1, W2,.., W L}, with W l being the node weight vector of the l-th layer, and l being a positive integer from 1 to L; Step 2: Cluster the weights of each layer of nodes in the deep neural network M. Step 3: Based on the clustering results, mark the redundant nodes in each layer of the deep neural network M. Step 4: Construct a network reliability calculation model based on the fault injection strategy, with the node failure probability P and the reliability R of M under P M Perform fault injection on non-redundant nodes and judge the redundant reliability of the deep neural network M based on the obtained reliability values.

2. The redundancy reliability measurement method for a deep neural network according to claim 1, characterized in that: The said Step 2 includes the following steps: Step 2.1: Based on the probability distribution of layer node weights, use the Gaussian kernel function to obtain the number of clusters k of the nodes W in the l-th layer l ; Step 2.1: Use the K-means algorithm to cluster the W of the nodes in the l-th layer l into k clusters, denoted as W l,1 , W l,2 ,.., W l,k .

3. A method for measuring the redundancy reliability of a deep neural network according to claim 2, characterized in that: The said Step 3 includes the following steps: Step 3.1: Initialize the loop variable j = 1. Step 3.2: Identify the geometric center of W as denoted by Equation (1) l,j and denote it as W GM , The function argmin returns the element that minimizes the sum of the Euclidean distances of the n elements in W l,j , where is the i-th element in W l,j , d is the dimension of , R d is a real number of dimension d, n is the number of elements in W l,j , and ||*||2 refers to the 2-norm; Step 3.3: Obtain the first found distance W according to formula (2). GM The element with the smallest Euclidean distance Identify W with this l,j The corresponding redundant network node; Step 3.4: According to the preset redundancy S, use Equation (2) to mark S×len(W l,j in l,j ) weights corresponding to the redundant nodes, and place them into the set W l,j-redundance , with W l,j-redundance being the weight set corresponding to the redundant nodes in the jth classification of the lth layer; Step 3.5: If j < k, then j = j + 1, and return to Step 3.2, otherwise, proceed to the next step. Step 3.6: Set the values of all elements in the weight set W l-redundance corresponding to the redundant nodes in the l-th layer to 0, and W l-redundance = {0}; Step 3.7: Mark the weights corresponding to the non-redundant nodes in the $l$-th layer with Equation (3), and place them into the set $W$. l-key ; W l-key = W l -W l-redundance (3) Step 3.8: Denote the updated model of M as M updated .

4. A method for measuring the redundant reliability of a deep neural network according to claim 3, characterized in that: The said Step 4 includes the following steps: Step 4.1: Initialize the loop variable E = 0. Step 4.2: Randomly extract a sample x from the sample data set V, extract the attribution probability vector vector1 of x based on M from the output end of softmax, and give the ranking rank_golden of the probability of the category of x through vector1; where softmax is the normalized exponential function. Step 4.3: Model M according to the given node failure probability P updated The set W l-key Implement fault injection and extract x from the output of softmax based on M updated The belonging probability vector vector2 is then used to give the probability ranking rank_fault for the category of x; Step 4.4: Take rank_golden1 and rank_fault1 as the first elements corresponding to rank_golden and rank_fault. If rank_golden1 = rank_fault1, then proceed to the next step, otherwise, return to Step 4.

2. Step 4.5: Calculate the output reliability R1 and cosine similarity R2 of M updated under the input x according to Equations (4) and (5). Among them, K is M updated The number of classifications of the output layer, and respectively represent the i-th element extracted from rank_golden and rank_fault, vector1 i and vector2 i respectively represent the i-th element extracted from vector1 and vector2, ||*||1 is the 1-norm; Step 4.6: E = E + 1, calculate the reliability R of M under the input x according to Equation (6) updated of M x , R x = αR1 + βR2 (6) Where α and β are the importance coefficients of R1 and R2 respectively, and α ∈ [0, 1], β ∈ [0, 1]. Step 4.7: If E > 1, then use formula (7) to obtain the unbiased estimate of M updated and proceed to the next step; otherwise, return to Step 4.2; where R′ represents the average value of the reliability of M under randomly sampling E times from V updated ; Step 4.8: Obtain the redundant reliability R of M according to Equation (8). R = R M -R′ (8) Step 4.9: If R satisfies convergence or E≥E max , stop the iteration and proceed to the next step; otherwise, return to Step 4.2; Step 4.10: Output the redundant reliability R of M under the redundancy S.

5. A method for measuring the redundancy reliability of a deep neural network according to claim 4, characterized in that: In the said Step 4.9, Equation (9) is satisfied for convergence. Among them, ε is a predefined convergence parameter and ε ∈ [0.01, 0.05].

6. A computer-readable storage medium, characterized in that, It stores a deep neural network redundant reliability measurement program, and when the deep neural network redundant reliability measurement program is executed by a processor, it implements the deep neural network redundant reliability measurement method as described in any one of Claims 1-5.

7. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the deep neural network redundant reliability measurement method as described in any one of Claims 1-5.

Images