Improved parallel channel convolutional neural network training method

Abstract

The invention relates to an improved parallel channel convolutional neural network training method. The improved parallel channel convolutional neural network training method comprises steps that characteristic extraction of data of the convolutional neural network is carried out through utilizing direct connection and convolutional channels to acquire characteristic matrixes; the characteristic matrixes are merged, and data dimension reduction is further carried out; the convolutional neural network is trained, and a loss value of network training at present is calculated; error items and a weight gradient of each layer are calculated; whether the network is in a convergence state is determined according to the loss value, if not, an initialization parameter of the convolutional neural network is adjusted according to the weight gradient, and re-training is further carried out; if yes, the network training result is outputted. The method is advantaged in that data circulation in the network can be guaranteed through introducing the direct connection channel, a problem of gradient instability during deep convolutional neural network training is solved, and deeper networks can be trained; through maximum pooling and mean value pooling, characteristic matrix dimensions of two times of characteristic extraction can be made to be consistent, and advantages of two pooling methods are integrated.

Description

technical field [0001] The invention belongs to the technical field of deep learning and big data, and in particular relates to an improved parallel channel convolutional neural network training method. Background technique [0002] With the development of society and the advent of the era of big data, related technologies continue to develop and innovate. Deep learning has made a series of breakthroughs in recent years because it can use massive data and improve classification accuracy through deeper network training. Scholars have tried to improve the performance of the convolutional neural network by increasing its size, and the easiest way to increase the network size is to increase the "depth". [0003] However, the deep network built based on the structure of the traditional convolutional neural network, as the number of network layers increases, the accuracy will reach saturation or even decrease. A multi-stage training method is proposed in the document "Romero A, ...

AI Technical Summary

Problems solved by technology

Doing so requires artificially adjusting multiple network parameters separately, which is time-consuming and labor-intensive, and training multiple shallow networks separately will lose the correlation information between the networks, which will affect the final performance of the network

The document "Lee C Y, Xie S, GallagherP, et al.Deeply-Supervised Nets[C] / / Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics.2015:562-570" introduces A number of auxiliary classifiers are introduced. Although this method can compensate the gradient disappearance problem when the deep network error is reversed to a certain extent, the introduction of auxiliary classifiers will also affect the final accuracy of the network.

Images