Pruning algorithm of convolutional neural network based on optimal threshold

Abstract

In the process of pruning, in order to automatically obtain an optimal pruning threshold that can balance the maximum sparse rate and the minimum error. This paper proposes a convolutional neural network pruning algorithm based on the optimal threshold. The algorithm uses the optimization ability of the greedy algorithm to select an optimal threshold, and uses the sensitivity and correlation of the node as factors to determine whether the node is important. Then by deleting the nodes whose importance is below the optimal threshold, the purpose of pruning the network is achieved. Experiments show that under the premise of loss accuracy within 2%, the algorithm can test the Lenet-5 network pruning on the M-NIST data set, which can accelerate 36.62%. This algorithm tests the VggNet network pruning on the CIFAR-10 dataset, which can speed up 43.86%. Experiments show that the algorithm effectively reduces network parameters and reduces running time.