Research On Multi-Objective Convolutional Neural Network Search Method Based On QUATRE Algorithm

Convolutional neural network,which plays an important role in the development of the whole society,can efficiently achieve image classification and can be used in the field of defect detection for industrial products and medical diagnosis.The traditional design for convolutional neural network mainly relies on manual design.However,artificial design of convolutional neural network requires a lot of experience and a lot of time.Using network model search method instead of traditional design for convolutional neural network can effectively make up for the above shortcomings.In addition,embedded processors which have the characteristics of small storage space and low computing power are widely used in the industrial field.However,most of the current network model search methods have problems such as high requirements for computational resource,large searched network models with high computational complexity.Aiming at these problems,this thesis proposes a multi-objective convolutional neural network model search method(SNQ-Net)based on QUANTRE algorithm.The SNQ-Net search method considers searching a convolutional neural network model with small computational complexity and high accuracy as a multi-objective optimization problem.In order to search for a convolutional neural network model with small computational complexity and high accuracy,SNQ-Net's search strategy uses an improved binary multi-objective QUATRE optimization algorithm.The binary multi-objective QUATRE optimization algorithm uses an adaptive flip matrix to flip the globally optimal population and alternately uses the globally optimal population of accuracy and the globally optimal population of computational complexity,so that the binary QUATRE algorithm can effectively balance the exploration and exploitation,searching the target network model;SNQ-Net evaluation strategy uses the proposed evaluation strategy based on network relative performance prediction.Proposed evaluation strategy predicts the relative performance of the two networks,sorts performance of all predicted convolutional neural networks by counting the relative performance for every convolutional neural network and selects networks with excellent performance to train,avoiding a lot of training of low-performance networks.In order to avoid repetitive training of the network model with the same structure and further reduce the requirements for computing resources,SNQ-Net adopts the proposed removal algorithm for duplicate encoding,based on the degree of access of nodes.The proposed algorithm gets the degree of access of all nodes by the encoding of the convolutional neural network,sorts the sequence consisting of the degree of access of all nodes and determines whether the corresponding convolutional neural network is trained by checking whether the sequence is in a dictionary tree.The experiment used the CIFAR-10 dataset.The experimental results show that compared with other search methods,SNQ-Net reduces the training of 95.6% of the networks.The entire search process only takes 0.8 GPU days,and the number of parameters in the searched convolutional neural network model is 2.7M,and the FLOPS value is 1024 M,which is smaller than most other convolutional neural networks.The accuracy on the CIFAR-10 test set reaches 95.72%,which exceeds most other searched convolutional neural network by comparative search methods.