An Information Theory-Inspired Strategy for Automated Network Pruning

Abstract

Despite superior performance achieved on many computer vision tasks, deep neural networks demand high computing power and memory footprint. Most existing network pruning methods require laborious human efforts and prohibitive computation resources, especially when the constraints are changed. This practically limits the application of model compression when the model needs to be deployed on a wide range of devices. Besides, existing methods are still challenged by the missing theoretical guidance, which lacks influence on the generalization error. In this paper we propose an information theory-inspired strategy for automated network pruning. The principle behind our method is the information bottleneck theory. Concretely, we introduce a new theorem to illustrate that the hidden representation should compress information with each other to achieve a better generalization. In this way, we further introduce the normalized Hilbert-Schmidt Independence Criterion on network activations as a stable and generalized indicator to construct layer importance. When a certain resource constraint is given, we integrate the HSIC indicator with the constraint to transform the architecture search problem into a linear programming problem with quadratic constraints. Such a problem is easily solved by a convex optimization method within a few seconds. We also provide rigorous proof to reveal that optimizing the normalized HSIC simultaneously minimizes the mutual information between different layers. Without any search process, our method achieves better compression trade-offs compared to the state-of-the-art compression algorithms. For instance, on ResNet-50, we achieve a 45.3%-FLOPs reduction, with a 75.75 top-1 accuracy on ImageNet. Codes are available at https://github.com/MAC-AutoML/ITPruner.