Pruning the deep neural network by similar function

Abstract

Recent deep neural networks become deeper and deeper, while the demand for low computational cost model will be higher and higher. The exists pruning algorithm usually focus on pruning the network layer by layer, or using the weight sum as important score. However, these methods do not work very well. In this paper, we propose a unified framework to accelerate and compress cumbersome CNN models. We put it into an optimization problem to find a subset of the model which can produce the most comparable outputs. We concentrate on filter level pruning. Experiment shows that our method has surpassed the exists filter level pruning algorithm. Taking the network as a whole is better than pruning it layer by layer. We also have an experiment on the large scale ImageNet dataset. The result shows that we can accelerate the VGG-16 by 3.18× without accuracy drop.