Work-in-Progress: A SIMD-Aware Pruning Technique for Convolutional Neural Networks with Multi-Sparsity Levels

Abstract

Designs of electronic systems often require considering multiple design concerns. In this paper, we propose a novel multi-phase pruning technique for convolutional neural networks (CNNs) that is capable of efficient exploration of multiple design objectives and constraints. To truly take advantage of the sparsity obtained by pruning, we present two different levels of pruning granularity, fine- and coarse-grain, and show how they are combined in the design space exploration. In particular, we propose to take the SIMD architecture into account in the fine-grain pruning. By iteratively pruning to a single CNN, multiple candidates can be obtained from the trade-off between the given design concerns. Experiments with existing CNNs verify that the proposed technique enables more efficient design space exploration over the accuracy-speed trade-off.