Robust Neural Network Pruning by Cooperative Coevolution

Abstract

Convolutional neural networks have achieved success in various tasks, but often lack compactness and robustness, which are, however, required under resource-constrained and safety-critical environments. Previous works mainly focused on enhancing either compactness or robustness of neural networks, such as network pruning and adversarial training. Robust neural network pruning aims to reduce computational cost while preserving both accuracy and robustness of a network. Existing robust pruning works usually require expert experiences and trial-and-error to design proper pruning criteria or auxiliary modules, limiting their applications. Meanwhile, evolutionary algorithms (EAs) have been used to prune neural networks automatically, achieving impressive results but without considering the robustness. In this paper, we propose a novel robust pruning method CCRP by cooperative coevolution. Specifically, robust pruning is formulated as a three-objective optimization problem that optimizes accuracy, robustness and compactness simultaneously, and solved by a cooperative coevolution pruning framework, which prunes filters in each layer by EAs separately. The experiments on CIFAR-10 and SVHN show that CCRP can achieve comparable performance with state-of-the-art methods