Pruning Convolutional Neural Networks Via Filter Similarity Analysis.

Geng Lili,Niu Baoning

DOI: https://doi.org/10.1007/s10994-022-06193-w

IF: 5.414

2022-01-01

Machine Learning

Abstract:Deep learning has shown excellent performance in many fields, especially image recognition and retrieval in recent years. The performance of convolutional neural networks (CNNs) is particularly outstanding. CNNs, however, are usually computationally intensive, which hinders the deployment of CNNs in resource-limited devices. Methods of network compression, pruning methods in particularly, removing redundant structures of CNNs, can significantly reduce the computational complexity of CNNs. Most of the state-of-the-art pruning methods for CNNs, however, have two defects. (1) Filters, also called convolutional kernels that are matrices used to extract features in an image, are pruned by ranking their weight without considering the effects of their actual output, which results in the deletion of important filters and the difficulty in determining the pruning threshold on weight. (2) Filters are pruned either in the forward direction or in isolation, which are difficult to control the loss of accuracy. This paper proposes a novel pruning method called filter similarity analysis with backward pruning (FSABP). FSABP calculates the similarity coefficients of filters in each layer, and deletes the filters associated with small similarity coefficients. The smaller the coefficient the more similar the filters. Filters are pruned layer by layer in the backward direction starting from the last convolution layer, which can effectively control the loss of accuracy by avoiding early removal of the shallow convolution filters. Experiments on LENET, VGG-16 and ResNet-50 show that FSABP can reduce parameter redundancy at the cost of negligible loss of accuracy and even improve accuracy in some cases. The results on LENET also suggest that FSABP is applicable to both deep and shallow CNNs.