Multi-column Deep Neural Network Based On Particle Swarm Optimization

Abstract

Many architectures for computer vision use of Convolutional Neural Networks (CNN) are biologically inspired by the receptive fields in the visual cortex. A modified deep architecture Thresholding Convolution Neural Network (ThCNN) progresses in this paper. This method is simple and effective to; regularize features map in the early layers of Convolution Neural Network; Compute the optimal global threshold to determine the features that are passed to the next layers; improve the speed performance of the suggested structure. The present paper proposed that each feature map in the convolution layer is connected to half feature maps in the previous layer. Another tendency in current applications is to combine the results of multiple networks, the particle swarm optimization method (PSO) is used to determine the influence of each network in the combining model. The proposed method was evaluated using the MNIST datasets. The experimental results showed that the error rate has decreased (up to 26.98 percent). Furthermore, combining model has decreased the number of the networks with the percent of 20%, and decreased error rate to 78.26 % of error rate of the best single model and to 88.89% of error of the averaging model.