CN115759200A - A key path construction method for deep network based on iterative optimization of gate parameters - Google Patents

Abstract

The invention discloses a gate parameter iterative optimization-based deep network key path construction method, which comprises the following steps of: step one, acquiring a data set; step two, constructing a deep convolution neural network model; adding gate parameters to the trained deep convolutional neural network model; and step four, constructing a critical path model of the depth network based on the gate parameters. The method has the advantages that the structure is simple, the design is reasonable, the importance of each convolution kernel is calculated based on the gate parameters, the convolution kernel with high contribution degree is selected to obtain the depth network key path model, when the constructed depth network key path model is repaired, only the gate parameters are updated and optimized, other parameters of the depth network model are not updated, and the parameters and the structure of the depth network model are not changed; the repairing reduces the covariate deviation and has good use effect.

Description

A key path construction method for deep network based on iterative optimization of gate parameters

technical field

The invention belongs to the technical field of artificial intelligence, and in particular relates to a method for constructing a key path of a deep network based on iterative optimization of gate parameters.

Background technique

Deep learning technology has achieved many good results, but there are still the following problems: 1. The network structure of the deep learning convolutional network model is huge, which brings high storage space and computing resource consumption, making it difficult to implement it in various Hardware platform. The current mainstream network, such as VGG16, has more than 130 million parameters and occupies more than 500 MB of space. It needs more than 30 billion floating-point operations to complete an image recognition task. 2. The structure of the deep neural network is complex and huge, and the law of information transfer and evolution in the network is not clear, resulting in poor interpretability of the intelligent algorithm based on deep learning, and the mechanism of intelligent decision-making is unclear, and users cannot clearly understand the deep learning convolutional network. The process of decision-making within the model and the basis for decision-making bring potential risks to practical applications. Therefore, in recent years, research on the transparency, interpretability, and credibility of deep learning convolutional network models has gradually increased.

Aiming at the complex problem of deep learning network structure and information evolution law, this application studies the key path construction method of deep learning network, analyzes key neuron nodes and paths of deep network, quantifies neuron contribution and simplifies deep network, and then improves deep learning network Interpretability of structure mechanism of action.

Contents of the invention

The technical problem to be solved by the present invention is to provide a method for constructing a key path of a deep network based on iterative optimization of gate parameters in view of the deficiencies in the above-mentioned prior art. Degree, select the convolution kernel with a large contribution to obtain the critical path model of the deep network. When repairing the constructed critical path model of the deep network, only the gate parameters are updated and optimized, and other parameters of the deep network model are not updated, and the depth is not changed. The parameters and structure of the network model; the fix reduces the covariate shift and works well.

In order to solve the above-mentioned technical problems, the technical solution adopted in the present invention is: a method for constructing a critical path of a deep network based on iterative optimization of gate parameters, characterized in that: comprising the following steps:

Step 1. Obtain a data set: collect multiple images, obtain a data set W=(X, Y), where X represents the input data, Y represents the data label, and divide the data set W into a training set, a verification set and a test set in proportion;

Step 2. Build a deep convolutional neural network model: define the objective function of the deep convolutional neural network model, use the training set as the input of the deep convolutional neural network model, and solve the optimal parameters of the deep convolutional neural network model to complete the deep convolutional neural network model training;

Step 3. Add gate parameters to the trained deep convolutional neural network model:

Step 301, adding gate parameters φ _i,j to the convolution kernels of each layer of the trained deep convolutional neural network model, φ _i,j represents the value added to the jth convolution kernel of the i convolutional layer Gate parameter, the gate parameter φ _{i, j} is (1, C _{i, j} , 1, 1), C _{i, j} represents the number of image channels of the gate parameter φ _{i, j} , 1≤i≤n, n represents the network Number of layers, 1≤j≤m, m represents the number of convolution kernels of the i-th layer;

Step 302: Use the training set to retrain the deep convolutional neural network model with gate parameters φ _i,j , use the verification set to verify the deep convolutional neural network model with gate parameters φ _i,j , and obtain the retrained The deep convolutional neural network model and gate parameters φ _i,j ;

Step 4. Construct the critical path model of the deep network based on the gate parameters:

与真实值Y的差值，Ω表示除了φ_i,j以外的网络模型参数；Step 401. Calculate the importance of each convolution kernel based on the gate parameters: the computer calculates the retrained depth convolution according to the formula Θ(φ _i,j )=KL(ΔL(φ _i,j )||ΔL _Ω (0)) The importance degree Θ(φ _i,j ) of the jth convolution kernel in the i-th layer of the product neural network model, where KL( ) represents the KL divergence calculation, and ΔL(φ _i,j ) represents the re-training in step 3 The difference between the predicted value of the deep convolutional neural network model and the real value, ΔL _Ω (0) represents the predicted value of the deep convolutional neural network model trained in step 2

The difference from the real value Y, Ω represents the network model parameters other than φ _i,j ;

Step 402, use the threshold to zero the gate parameters with low importance of the convolution kernel: if Θ(φ _i,j )<Θ(φ), assign the gate parameters φ _i,j to 0, and Θ(φ) represents the optimized threshold , to obtain the critical path model of the deep network;

α表示权重,X_k表示训练集中的第k个数据，Y_k表示X_k对应的样本标签，1≤k≤N，N表示训练集的样本数，Δ表示标签平滑因子，p(X_k,θ^-)表示训练样本X_k的预测概率，θ^-表示关键路径模型的参数；Step 403. Repair the constructed deep network critical path model: input the training set into the deep network critical path model for iterative training, and update and optimize the gate parameters of the deep network critical path model through the gradient descent algorithm. The zeroed gate parameters are always Set to zero, do not participate in the update, the loss function of the deep network critical path model in the iterative training process of the gate parameters is L'=(1-α)L ₁ +αL ₂ , where

α represents the weight, X _k represents the kth data in the training set, Y _k represents the sample label corresponding to X _k , 1≤k≤N, N represents the number of samples in the training set, Δ represents the label smoothing factor, p(X _k , θ ^- ) represents the predicted probability of the training sample X _k , θ ^- represents the parameters of the critical path model;

Step 404: Use the verification set to verify the updated and optimized deep network critical path model in step 403 to obtain a repaired deep network critical path model;

Step 405, input the test set into the repaired deep network critical path model, and obtain the classification result of the test set.

The above-mentioned method for constructing a critical path of a deep network based on iterative optimization of gate parameters is characterized in that: in step 1, the image is a radar image, and the foreground image is first extracted from the collected radar image, and then Gaussian blur processing is performed, and then the feature is extracted Vector, get the data set W=(X,Y).

其中L(X，Y:θ)表示交叉熵损失函数，θ表示网络模型的参数，

表示门参数的稀疏约束，λ表示稀疏约束的权重。The above-mentioned method for constructing a critical path of a deep network based on iterative optimization of gate parameters is characterized in that: in step 302, the loss function of the network model of the gate parameters φ _i,j is added

Where L(X, Y:θ) represents the cross-entropy loss function, θ represents the parameters of the network model,

Denotes the sparse constraint on the gate parameters, and λ denotes the weight of the sparse constraint.

L表示损失函数，δL表示损失函数L的增量，δφ_i,j表示门参数φ_i,j的增量，

表示样本预测结果，Y表示样本标签。The above-mentioned method for constructing a critical path of a deep network based on iterative optimization of gate parameters is characterized in that: in step 401,

L represents the loss function, δL represents the increment of the loss function L, and δφ _i,j represents the increment of the gate parameter φ _i,j ,

Indicates the sample prediction result, and Y indicates the sample label.

Compared with the prior art, the present invention has the following advantages:

1. The structure of the present invention is simple, the design is reasonable, and the realization, use and operation are convenient.

2. The present invention adds gate parameters to the trained deep convolutional neural network model, calculates the importance of each convolution kernel based on the gate parameters, and obtains convolution kernel nodes with large contribution and strong information interaction, which can improve the deep network structure The interpretability of the mechanism of action, the selection of a convolution kernel with a large contribution can obtain the critical path model of the deep network, and play a role in simplifying the deep network.

3. When the present invention repairs the constructed deep network critical path model, only the gate parameters of the deep network critical path model are updated and optimized through the gradient descent algorithm. The zeroed gate parameters are always set to zero, do not participate in the update, and do not correct Other parameters of the deep network model are updated without changing the parameters and structure of the deep network model.

4. The present invention adopts the superposition of two loss functions to construct a new loss function when the network model is repaired and trained, and the loss function L ₁ adds the constraint conditions of the gate parameters φ _i,j , and the loss function L ₂ introduces a label smoothing factor, The label pair smoothing process reduces the influence of noise in the label, improves the accuracy of the label to data, further reduces the covariate offset, and improves the recognition accuracy of the network model.

To sum up, the present invention has a simple structure and a reasonable design. The importance of each convolution kernel is calculated based on the gate parameters, and the convolution kernel with a large contribution is selected to obtain the critical path model of the deep network. When repairing, only the gate parameters are updated and optimized, and other parameters of the deep network model are not updated, and the parameters and structure of the deep network model are not changed; the repair reduces the covariate offset, and the use effect is good.

The technical solutions of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments.

Description of drawings

Fig. 1 is the block diagram of circuit principle of the present invention.

Detailed ways

The method of the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments of the present invention.

It should be noted that, in the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and examples.

It should be noted that the terminology used here is only for describing specific implementations, and is not intended to limit the exemplary implementations according to the present application. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural, and it should also be understood that when the terms "comprising" and/or "comprising" are used in this specification, they mean There are features, steps, operations, means, components and/or combinations thereof.

It should be noted that the terms "first" and "second" in the description and claims of the present application and the above drawings are used to distinguish similar objects, but not necessarily used to describe a specific sequence or sequence. It is to be understood that the data so used are interchangeable under appropriate circumstances such that the embodiments of the application described herein, for example, can be practiced in sequences other than those illustrated or described herein. Furthermore, the terms "comprising" and "having", as well as any variations thereof, are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or device comprising a sequence of steps or elements is not necessarily limited to the expressly listed instead, may include other steps or elements not explicitly listed or inherent to the process, method, product or apparatus.

For the convenience of description, spatially relative terms may be used here, such as "on ...", "over ...", "on the surface of ...", "above", etc., to describe the The spatial positional relationship between one device or feature shown and other devices or features. It will be understood that the spatially relative terms are intended to encompass different orientations of the device in use or operation in addition to the orientation depicted in the figures. For example, if the device in the figures is turned over, devices described as "above" or "above" other devices or configurations would then be oriented "beneath" or "above" the other devices or configurations. under other devices or configurations”. Thus, the exemplary term "above" can encompass both an orientation of "above" and "beneath". The device may be otherwise oriented (rotated 90 degrees or at other orientations) and the spatially relative descriptions used herein interpreted accordingly.

As shown in Figure 1, the present invention comprises the following steps:

Step 1. Obtain a data set: collect multiple images to obtain a data set W=(X, Y), where X represents the input data and Y represents the data label, and the data set W is divided into training set, verification set and test set in proportion.

In actual use, the image is a radar image collected according to the application scene, or the MSTAR dataset is directly used. The MSTAR dataset is widely used in the research of radar image target recognition.

For the collected radar image, first extract the foreground image, then perform Gaussian blur processing, and then extract the feature vector, the complexity of the obtained feature vector becomes lower, removes the offset and unimportant features, reduces the complexity of the input information, and improves The distribution uniformity rate of the sample to the covariate improves the effect of the model.

The feature vectors are saved or stored on an electronic or computer readable storage medium and/or transmitted to another program or application for further processing or use.

Step 2. Build a deep convolutional neural network model: define the objective function of the deep convolutional neural network model, use the training set as the input of the deep convolutional neural network model, and solve the optimal parameters of the deep convolutional neural network model to complete the deep convolutional neural network Model training.

It should be noted that the convolutional neural network model can choose vgg model, resnet-34 model, resnet-50 or resnet-56 model. The following deep learning neural network models, deep neural network models, neural network models, network models, and deep networks all refer to deep convolutional neural network models.

Step 3. Add gate parameters to the trained deep convolutional neural network model:

Step 301, adding gate parameters φ _i,j to the convolution kernels of each layer of the trained deep convolutional neural network model, φ _i,j represents the value added to the jth convolution kernel of the i convolutional layer Gate parameter, the gate parameter φ _{i, j} is (1, C _i , 1, 1), C _i represents the number of image channels of the gate parameter φ _{i, j} , 1≤i≤n, n represents the number of network layers, 1 ≤j≤m, m represents the number of convolution kernels in the i-th layer.

The gate parameter is a gradient-updatable variable. The function of the convolutional layer is to extract the features of the image. The input data of the convolutional layer is called the input feature map, and the output data is called the output feature map. For the setting of the gate parameter, the output of the convolutional layer is multiplied by a gate parameter, as the following The input of a convolutional layer, which is used to identify the properties of the convolutional kernel, changes the structure of the standard neural network model. The parameter of the gate parameter φ _i,j is (1, C _i,j ,1,1), and C _i,j represents the number of image channels of the i-th gate parameter. In actual use, if it is for the global simplification of the neural network model, the gate parameter φ _i,j is added after each convolutional layer, and C _i,j ≠1. If it is for partial simplification of the neural network model, C _i,j =1 of the simplified convolutional layer is not required.

其中L(X，Y:θ)表示交叉熵损失函数，θ表示网络模型的参数，

表示门参数的稀疏约束，λ表示稀疏约束的权重。Step 302, use the training set to retrain the deep convolutional neural network model with gate parameters φ _i,j , in actual use, the loss function of the network model with gate parameters φ _i,j

Where L(X, Y:θ) represents the cross-entropy loss function, θ represents the parameters of the network model,

Denotes the sparse constraint on the gate parameters, and λ denotes the weight of the sparse constraint.

After the training is completed, the parameters of the convolution kernel are fixed, and the verification set is used to verify the deep convolutional neural network model with gate parameters φ _i,j, and the retrained deep convolutional neural network model and gate parameters φ _i,j are obtained.

In actual use, use the training set to perform sparse training on the deep convolutional neural network model with gate parameters φ _i,j . The sparse constraints will reduce the dependence between each convolution kernel and make it more independent, which is convenient for further explanation of convolution. contribution of the nucleus. During training, only the gate parameters φ _i,j are updated, and a subset of training data is used to train the neural network model each time.

Step 4. Construct the critical path model of the deep network based on the gate parameters:

与真实值Y的差值，Ω表示除了φ_i,j以外的网络模型参数。Step 401. Calculate the importance of each convolution kernel based on the gate parameters: the computer calculates the retrained depth convolution according to the formula Θ(φ _i,j )=KL(ΔL(φ _i,j )||ΔL _Ω (0)) The importance Θ(φ _i,j ) of the jth convolution kernel in the i-th layer of the product neural network model, where KL( ) represents the KL divergence calculation, and ΔL(φ _i,j ) represents the re-trained in step 3 The difference between the predicted value of the deep convolutional neural network model and the real value, ΔL _Ω (0) represents the predicted value of the deep convolutional neural network model trained in step 2

The difference from the true value Y, Ω represents the network model parameters other than φ _i,j .

In a deep convolutional neural network, the role of the convolution kernel is to extract the features of the input radar data. For the convolution kernel responsible for extracting a certain feature, if it detects its corresponding feature from the input of the previous layer, the node is activated, allowing useful feature information to propagate to the next layer. And if the node thinks that the input does not contain its corresponding features, the node will be shielded to block the propagation of useless feature information to the next layer. Therefore, calculating the contribution degree of each convolution kernel based on the gate parameters and obtaining convolution kernel nodes with large contribution and strong information interaction is conducive to realizing the interpretability of the deep network and realizing the construction of the critical path model.

L表示损失函数，δL表示损失函数L的增量，δφ_i,j表示门参数φ_i,j的增量，

表示样本预测结果，Y表示样本标签。Based on the gate parameters φ _i,j, the contribution degree of each convolution kernel of each layer of the deep network is calculated, and the contribution degree is used as the basis for judging whether the convolution kernel is a key node. In actual use,

L represents the loss function, δL represents the increment of the loss function L, and δφ _i,j represents the increment of the gate parameter φ _i,j ,

Indicates the sample prediction result, and Y indicates the sample label.

Step 402, use the threshold to zero the gate parameters with low importance of the convolution kernel: if Θ(φ _i,j )<Θ(φ), assign the gate parameters φ _i,j to 0, and Θ(φ) represents the optimized threshold , to get the critical path model. In actual use, set the optimization threshold Θ(φ), filter the importance of the convolution kernel, and the convolution kernel whose contribution is lower than the preset optimization threshold Θ(φ) can be regarded as an "uncritical" node , so the corresponding gate parameter φ _i,j is assigned a value of 0, and the gate parameter is assigned a value of 0, which is equivalent to deleting the corresponding convolution kernel. The remaining convolution kernel is the convolution kernel that plays a key role in processing image information, and the remaining convolution kernel constitutes the critical path model of the deep network. The critical path model can be used to analyze the evolution of information in the critical path of the deep network, thus completing the interpretable elaboration of the deep network.

α表示权重,X_k表示训练集中的第k个数据，Y_k表示X_k对应的样本标签，1≤k≤N，N表示训练集的样本数，Δ表示标签平滑因子，p(X_k,θ^-)表示训练样本X_k的预测概率，θ^-表示关键路径模型的参数。Step 403. Repair the constructed deep network critical path model: input the training set into the deep network critical path model for iterative training, and update and optimize the gate parameters of the deep network critical path model through the gradient descent algorithm. The zeroed gate parameters are always Set to zero, do not participate in the update, the loss function of the deep network critical path model in the iterative training process of the gate parameters is L'=(1-α)L ₁ +αL ₂ , where

α represents the weight, X _k represents the kth data in the training set, Y _k represents the sample label corresponding to X _k , 1≤k≤N, N represents the number of samples in the training set, Δ represents the label smoothing factor, p(X _k , θ ^- ) represents the predicted probability of the training sample X _k , θ ^- represents the parameters of the critical path model.

In actual use, the loss value of the iterative training is backpropagated, and the gradient of the weight of the critical path model of the deep network interpretability is extracted during the backpropagation process. According to the extracted gradient, only the critical path model of the deep network The gate parameters are updated and optimized, and other parameters of the neural network model are not updated.

It should be noted that the deletion of the convolution kernel of the original deep network model is likely to cause information loss, which will cause the covariate shift of the optimized neural network model, resulting in a poor model effect. Therefore, when repairing training, Consider the optimization of the gate parameters φ _i,j , so the constraints of the gate parameters φ _i,j are added to the loss function L ₁ .

The loss function L ₂ can measure the difference between the predicted probability and the real probability of X _k . In the network model, the difference between the real probability distribution and the predicted probability distribution is expressed as a loss. The smaller the loss, the better the prediction classification effect of the network model. Δ=0.05, the label pairs are smoothed, which reduces the influence of noise in the labels, improves the correct rate of labels to data, further reduces the covariate shift, and improves the recognition accuracy of the network model.

Step 404: Use the verification set to verify the updated and optimized deep network critical path model in step 403 to obtain a repaired deep network critical path model;

Step 405, input the test set into the repaired deep network critical path model, and obtain the classification result of the test set.

Effectiveness of the present invention can further confirm by following simulation experiments:

1. Experimental conditions and methods

The hardware platform is: Inter(R)Core(TM)i5-10600K CPU@4.10GHZ,

16.0GB RAM;

The software platform is: Pytorch 1.10;

2. Simulation content and results

In this embodiment, the deep network selects the VGG16 neural network model. The VGG16 neural network model can be used for image processing, including the processing of radar image information. Image processing is used in a variety of technical applications. As an example, it can be used for the location of specific types or types of objects in images, such as in In computer vision applications for image and video analysis, this embodiment is not specifically limited.

Table 1 is a comparison table of the number of convolution kernel nodes and classification accuracy of each layer of the original deep network and the critical path model of the deep network when the deep network is divided into the VGG16 model.

Table 1 Comparison table between the original deep network and the critical path model of the deep network

For example, in the first layer, the number of original convolution kernels is 64, and only 51 convolution kernels that contribute to network performance are left after the contribution degree screening. It can be seen from the results that the deep network critical path model only retains 1452 convolution kernels of the original neural network model, that is, only 34.37% of the number of nodes is retained. After the simplification and repair, the correct recognition rate of the deep network critical path model on the data set is still 92.46%, and the performance of the critical path model is very close to the original deep network model, which shows the reliability of the deep network critical path model and verifies the The validity of the application's build method.

Moreover, from the changes in the number of convolution kernel nodes in each layer, it can be seen that no matter it is the deep network critical path model or the original deep network model, the number of convolution kernel nodes in the first layer is large, which is convenient for extracting enough information from the input samples. It is used for classification; the number of convolution kernel nodes in the last layer is also large, in order to provide enough information for the final classification. In contrast, the number of convolution kernel nodes in the middle layer is relatively small, indicating that during the transfer of features from input to output, the contribution of middle layer nodes to information transfer and extraction is low, and the neural network is performing task discrimination. The high-level and abstract features required by the system are less, so there is a lot of redundancy in the middle layer nodes. Using the key path can explain the evolution and transmission process of the signal in the network model, thereby improving the interpretability of the mechanism of the deep learning network structure.

It should be understood that the sequence numbers of the steps in the above embodiments do not mean the order of execution, and the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiment of the present application.

The above is only an embodiment of the present invention, and does not limit the present invention in any way. Any simple modifications, changes and equivalent structural changes made to the above embodiments according to the technical essence of the present invention still belong to the technical solution of the present invention. within the scope of protection.

Claims (4)

1. A deep network key path construction method based on gate parameter iterative optimization is characterized by comprising the following steps: the method comprises the following steps:

step one, acquiring a data set: acquiring a plurality of images to obtain a data set W = (X, Y), wherein X represents input data, Y represents a data label, and the data set W is divided into a training set, a verification set and a test set in proportion;

step two, constructing a deep convolution neural network model: defining a target function of a deep convolutional neural network model, inputting a training set as the deep convolutional neural network model, and solving the optimal parameters of the deep convolutional neural network model so as to complete the training of the deep convolutional neural network model;

adding gate parameters to the trained deep convolutional neural network model:

301, respectively adding a gate parameter phi to the convolution kernels of each layer of the trained deep convolution neural network model _i,j ，φ _i,j Representing the gate parameter added to the ith convolution layer and the jth convolution kernel, said gate parameter phi _i,j Is (1,C) _i,j ，1，1),C _i,j Representing the door parameter phi _i,j I is more than or equal to 1 and less than or equal to n, n represents the number of network layers, j is more than or equal to 1 and less than or equal to m, and m represents the number of convolution kernels of the ith layer;

step 302, add door parameter phi by training set pair _i,j The deep convolutional neural network model is trained again, and the verification set pair is used for adding the gate parameter phi _i,j The deep convolutional neural network model is verified to obtain a retrained deep convolutional neural network model and a gate parameter phi _i,j ；

Step four, constructing a critical path model of the depth network based on the gate parameters:

step 401, calculating the importance of each convolution kernel based on the gate parameters: the computer follows the formula theta (phi) _i,j )＝KL(ΔL(φ _i,j )||ΔL _Ω (0) Calculates the importance theta (phi) of the jth convolution kernel of the ith layer network of the retrained deep convolution neural network model _i,j ) Where KL (. Cndot.) represents the KL divergence calculation, Δ L (. Phi.) _i,j ) The difference value of the predicted value and the true value of the depth convolution neural network model retrained in the third step, delta L _Ω (0) Representing the predicted value of the deep convolutional neural network model which completes the training in the second step

Difference from the true value Y, Ω represents the difference except φ _i,j (ii) network model parameters;

step 402, zeroing the gate parameter with low importance of the convolution kernel by using a threshold value: if theta (phi) _i,j ) < theta (phi), the gate parameter phi is set _i,j The value is assigned to be 0, and theta (phi) represents an optimization threshold value to obtain a deep network key path model;

step 403, repairing the built deep network critical path model: inputting a training set into a deep network key path model for iterative training, updating and optimizing gate parameters of the deep network key path model through a gradient descent algorithm, setting the zero gate parameters to zero all the time without participating in updating, wherein the loss function of the deep network key path model in the gate parameter iterative training process is L' = (1-alpha) L ₁ +αL ₂ Wherein

Alpha represents a weight, X _k Representing the kth data in the training set, Y _k Represents X _k Corresponding sample label, k is more than or equal to 1 and less than or equal to N, N represents the number of samples of the training set, delta represents the label smoothing factor, and p (X) _k ,θ ^- ) Representing training sample X _k Predicted probability of (a), theta ^- Parameters representing a critical path model;

step 404, verifying the deep network key path model which is updated and optimized in the step 403 by using a verification set to obtain a repaired deep network key path model;

and 405, inputting the test set into the repaired deep network key path model to obtain a classification result of the test set.

2. The method for constructing the deep network critical path based on the gate parameter iterative optimization as claimed in claim 1, wherein: in the first step, the image is a radar image, a foreground image is firstly extracted from the acquired radar image, then Gaussian blur processing is performed, and then a feature vector is extracted, so that a data set W = (X, Y) is obtained.

3. The method for constructing the deep network critical path based on the gate parameter iterative optimization as claimed in claim 1, wherein: in step 302, a gate parameter φ is added _i,j Loss function of network model of

Where L (X, Y: θ) represents a cross entropy loss function, θ represents a parameter of the network model,

represents the sparse constraint of the gate parameters and λ represents the weight of the sparse constraint.

4. A door parameter based on as in claim 1The iterative optimization deep network key path construction method is characterized by comprising the following steps: in a step 401, in which the data is processed,

l denotes the loss function, δ L denotes the increment of the loss function L, δ φ _i,j Representing the door parameter phi _i,j The increment of (a) is increased by (b),

denotes the sample prediction result, and Y denotes the sample label.

Images