Research On Deep Learning Optimization Method Based On Genetic Algorithm

In recent years,deep learning has become popular in various fields due to its excellent performance.Deep learning network models represented by convolutional neural networks,residual neural networks and densely connected networks have achieved good recognition results in the field of image recognition.In traditional deep learning,deep networks often contain more semantic information.The study found that within a certain range,the classification ability of deep learning is proportional to the number of network layers,but when the network reaches a certain depth,the convergence speed of the network starts to slow down or even stops,and the classification ability of the network no longer rises or even declines trend.When the number of network layers is deepened,the number of network training parameters continues to increase,causing the gradient information responsible for network update to decay at a huge speed,making it difficult to update the parameters of neurons.As a result,problems such as the disappearance of the gradient,the failure of the network to converge,and the dispersion of the gradient appear.Moreover,the number of deep network layers and the large number of training parameters require researchers to spend a lot of time to determine the optimal structure and parameters to obtain better recognition accuracy.To solve the above problem,this dissertation carried out the following:On the one hand,this dissertation optimizes the ResNet through Genetic Algorithm and proposes GA-ResNet.In the GA-ResNet network,this dissertation defines the concept of stage.The stage is bounded by the pooling layer and consists of a series of predefined residual blocks.At the same time,this dissertation proposes a new coding rule GARCode,which converts the network structure into a fixed-length binary string.After the selection,mutation and crossover operation of the genetic algorithm,a new competitive network structure is generated from the previous generation population and eliminate weak network structure.The network uses the unique jump connection structure of the residual neural network to solve the problems of gradient dispersion and gradient disappearance that occur during network training.The powerful search ability of genetic algorithm can automatically find the optimal network structure and training parameters,which effectively improves the efficiency and accuracy of network classification.On the other hand,this dissertation optimizes the Dense Net through genetic algorithm and proposes GA-Dense Net.GA-Dense Net divides different stages with the transition layer as the boundary,and at the same time uses the dense block as the node,and uses the proposed new coding rule GADCode to perform genetic operations.Compared with GA-ResNet,the GA-Dense Net network structure is more complicated.The Bernoulli distribution is used to sample each individual independently in the initialization phase,and the "Russian Roulette" program is used to determine the surviving individuals in the individual selection phase,mutation and crossover operations are used to avoid the network falls into a local optimum,and finally the fitness function is used to evaluate the performance of the network structure.Dense Net can directly integrate shallow features with deep features,effectively improving the utilization of network features.GA-Dense Net combines the advantages of genetic algorithm and Dense Net,which can reduce training parameters,reduce system resource consumption,and improve the network's optimization ability.On the Cifar10 data set,the GA-ResNet network and the GA-Dense Net network achieved 91.26% and 91.84% classification accuracy,which are 2.43% and 3.01% higher than CNN;on the Cifar100 data set,the two networks achieved 65.13%,68.53%classification accuracy,which is 2.55% and 5.95% higher than CNN.The experimental results show that the GA-ResNet and the GA-Dense Net proposed in this dissertation have better performance than the traditional deep learning neural network.