Dynamical Channel Pruning by Conditional Accuracy Change for Deep Neural Networks

• Published in: IEEE transaction on neural networks and learning systems 2021-02, Vol.32 (2), p.799-813
• Main Authors: Chen, Zhiqiang, Xu, Ting-Bing, Du, Changde, Liu, Cheng-Lin, He, Huiguang
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Artificial Intelligence; Artificial neural networks; Automation; Biological neural networks; Channel estimation; Channels; Compression; Computer architecture; Computer Systems; Conditional accuracy change (CAC); Convolution; Data Compression; Databases, Factual; Deep Learning; Design criteria; direct criterion; dynamical channel pruning; Humans; Image Processing, Computer-Assisted - methods; Iterative methods; Logic gates; neural network compression; Neural networks; Neural Networks, Computer; Pattern Recognition, Automated; Pruning; structure shaping; Training; Weight
• ISSN: 2162-237X; 2162-2388
• DOI: 10.1109/TNNLS.2020.2979517

Channel pruning is an effective technique that has been widely applied to deep neural network compression. However, many existing methods prune from a pretrained model, thus resulting in repetitious pruning and fine-tuning processes. In this article, we propose a dynamical channel pruning method, which prunes unimportant channels at the early stage of training. Rather than utilizing some indirect criteria (e.g., weight norm, absolute weight sum, and reconstruction error) to guide connection or channel pruning, we design criteria directly related to the final accuracy of a network to evaluate the importance of each channel. Specifically, a channelwise gate is designed to randomly enable or disable each channel so that the conditional accuracy changes (CACs) can be estimated under the condition of each channel disabled. Practically, we construct two effective and efficient criteria to dynamically estimate CAC at each iteration of training; thus, unimportant channels can be gradually pruned during the training process. Finally, extensive experiments on multiple data sets (i.e., ImageNet, CIFAR, and MNIST) with various networks (i.e., ResNet, VGG, and MLP) demonstrate that the proposed method effectively reduces the parameters and computations of baseline network while yielding the higher or competitive accuracy. Interestingly, if we Double the initial Channels and then Prune Half (DCPH) of them to baseline's counterpart, it can enjoy a remarkable performance improvement by shaping a more desirable structure.