A Channel Merging Approach to Control Sparsity in Neural Networks

Abstract

본 논문은 가중치 밀집도가 낮은 신경망에서의 가중치 스케줄 효율을 높이기 위해 1) 가중치 채널 병합(channel-merging)을 통해 밀집도가 높은 신경망을 만들어 스케줄 효율을 높이고, 2) 병합된 가중치 채널을 다룰 수 있는 하드웨어 가속기를 제안 및 평가한다.

In this thesis, a channel-merging offline scheduling scheme is presented for improving the efficiency of the previous offline scheduler in highly pruned convolutional neural networks (CNN). In the channel-merging step, two channels in the same layers are merged lane-wise to increase the network’s channel-level sparsity. Also, a modified hardware architecture is presented to handle merged and scheduled weights. With the zero-skip and outlier-aware scheduling schemes of the previous accelerator, the proposed merging and scheduling method achieve more lane utilization and speedup. Despite a little area overhead of the proposed hardware, fast calculation and reduced memory access make the energy consumption lower than the previous hardware.