Pruning method using correlation of weight changes and weight magnitudes in CNN

초록

Very complex deep learning models need to be compressed to be memory and cost effective, especially for applications on a mobile platform. We propose a new method of selecting weights to prune to compress convolutional neural networks. To select unimportant weights and get the best result, we combine typical weight magnitude pruning method with our method, which evaluates correlation coefficients of weights to measure the strength of a relationship between weight magnitudes and weight changes through the iterations. In the experimental section, we show our result of pruning 94% of weights in LeNet-5 without significant accuracy loss. © The Korean Institute of Intelligent Systems.

키워드