EasiEdge: A Novel Global Deep Neural Networks Pruning Method for Efficient Edge Computing

Abstract

Deep neural networks (DNNs) have shown tremendous success in many areas, such as signal processing, computer vision, and artificial intelligence. However, the DNNs require intensive computation resources, hindering their practical applications on the edge devices with limited storage and computation resources. Filter pruning has been recognized as a useful technique to compress and accelerate the DNNs, but most existing works tend to prune filters in a layerwise manner, facing some significant drawbacks. First, the layerwise pruning methods require prohibitive computation for per-layer sensitivity analysis. Second, layerwise pruning suffers from the accumulation of pruning errors, leading to performance degradation of pruned networks. To address these challenges, we propose a novel global pruning method, namely, EasiEdge, to compress and accelerate the DNNs for efficient edge computing. More specifically, we introduce an alternating direction method of multipliers (ADMMs) to formulate the pruning problem as a performance improving subproblem and a global pruning subproblem. In the global pruning subproblem, we propose to use information gain (IG) to quantify the impact of filters removal on the class probability distributions of network output. Besides, we propose a Taylor-based approximate algorithm (TBAA) to efficiently calculate the IG of filters. Extensive experiments on three data sets and two edge computing platforms verify that our proposed EasiEdge can efficiently accelerate DNNs on edge computing platforms with nearly negligible accuracy loss. For example, when EasiEdge prunes 80% filters in VGG-16, the accuracy drops by 0.22%, but inference latency on CPU of Jetson TX2 decreases from 76.85 to 8.01 ms.