Research On Convolutional Neural Network Model Compression Technology Based On Genetic Algorithm

As a leader in the field of artificial intelligence,deep learning technology has made great progress in the past decade,and has become an irreplaceable mainstream algorithm in many fields such as computer vision and natural language processing.With the increasing complexity of the convolutional neural network structure,model compression technology has begun to receive more and more attention from researchers.Network pruning,as an important research direction of model compression technology,has achieved remarkable results in recent years.However,the investigation in this paper found that there are still few researchers trying to combine network pruning technology with evolutionary algorithms to perform model compression.This is because evolutionary algorithms are difficult to embed in convolutional neural networks,and there are problems that subsequent models are difficult to optimize.Based on the above discussion,in order to expand the research ideas of the network pruning technology and supplement the automatic model compression method,this paper considers combining the network pruning technology with the currently mature evolutionary algorithm-genetic algorithm,and developed a set of genetic algorithm based convolutional neural network model compression technology.This technology uses genetic algorithms to optimize the pruning of the network model,and combines the back propagation algorithm to optimize the model parameters.The process of genetic iteration can effectively prevent the back propagation from falling into the local optimum,thereby maintaining the accuracy of the model,and greatly compressing and accelerating the convolutional neural network.The compressed model has a larger reduction in the model size,calculation cost and other indicators than the original model,and the algorithm does not require a specific calculation library support.Experiments prove that the method proposed in this paper can meet the actual needs,achieve a good model compression effect,and at the same time reduce the calculation amount of the model on a large scale,increase the response rate of the model.The innovations of this paper are as follows:1.A convolutional neural network model branching method based on combined channel pruning is proposed.This method implements branching processing of the network model through multiple combined channel pruning methods and model fine-tuning to form a multi-branch network structure.So as to reduce the storage scale and calculation cost of the model while maintaining the accuracy of the model.2.Based on the multi-branch neural network model,a pruning technology of multi-branch neural network based on genetic algorithm is proposed.This technology uses the genetic algorithm's mechanism of survival of the fittest to gradually gather superior branches in the multi-branch neural network,eliminate the inferior branches,and supplement the model fine-tuning.In this way,the network model is further effectively compressed.