A Novel Approach to Structured Pruning of Neural Network for Designing Compact Audio-Visual Wake Word Spotting System
Haotian Wang, Jun Du, Hengshun Zhou, Heng Lu, Yuhang Cao · 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) · 2022
In this paper, we propose a novel approach to structured pruning of neural network. Firstly, we extend the original channel-level pruning from one-shot manner to iterative manner. Then we further employ the learning rate rewinding strategy in the lottery ticket hypothesis (LTH) to guide the channel-level pruning, yielding a new algorithm named channel-level pruning with learning rate rewinding (CPLR). Finally, we apply CPLR to prune the audio and video networks for designing compact audio-visual wake word spotting (AVWWS) system. Tested on MISP-2021 AVWWS database, the results show that the proposed CPLR approach performs better than either the channel-level pruning approach or LTH approach in term of both system performance and model efficiency. More interestingly, we observe that while the network parameters are greatly reduced by CPLR, the network generalization capability can be even better.