An Efficient Convolutional Neural Network With Attached Accelerating Strategy

Abstract

With the development of convolution neural network, people invented a deeper network and gained higher accuracy. But with the increasing demand in mobile technology of this field, how to save the computing resource consumption, enhance the training speed has become an important question. In this paper, based on some of the classic acceleration strategy, we established a new multi-branch network called FRINet, the model lowers computing power of the hardware requirements and enhancing training efficiency. By experiments on ISIC dataset, as compared to no acceleration strategy model InceptionV3, we achieve 2.3 times speedup in training speed, while the loss of the correct rate is only 1.7%.