Convolutional neural network simplification via feature map pruning

Abstract

Convolutional neural networks (CNNs) have been a focus area of machine learning in recent years, and they are widely used in vision and speech processing because of their superior performance. However, CNNs are usually resource-heavy to ensure higher accuracy, i.e., an accurate network with millions of parameters requires high performance computing devices. This prevents the use of CNNs in resource-limited hardware. In this paper, we propose a novel CNN simplification method to prune feature maps with relatively low discriminability magnitudes, which can produce a simplified CNN with reduced computational cost. Specifically, we define the critical points among the discriminability values of feature maps in each convolutional layer, and use these critical points to easily find the best pruning number of feature maps. Our experimental results show that in each convolutional layer of the VGG model, 15.6% to 59.7% of feature maps can be pruned without any loss of accuracy in classification tasks.