报告主题：Constructing Spiking Neural Networks from Pretrained Artificial Neural Networks

报告人：Gu, Shi (顾实)

报告人单位：University of Electronic Science and Technology of China

主办：数字媒体研究所

时间：4月23日（周五）上午10点

地点：理科二号楼2735

讲座摘要

Spiking Neural Network (SNN) has been recognized as one of the next-generation neural networks. Despite the success of SNNs in several applications, it remains a challenging task to convert a pretrained Artificial Neural Network (ANN) to a high-performance SNN. In this talk, we first give some theoretical analyses of the decomposition of the conversion error and how it propagates through layers. Based on the analyses, we propose an adaptive method to adjust the threshold, potential, and weight with respect to the simulation length to improve the firing rate. Empirically, we demonstrate our method can successfully handle the SNN conversion with Batch Normalization layers for the first time and effectively preserve the high accuracy even in just a few time steps. For example, our advanced pipeline can increase up to 69% top-1 accuracy when converting Mobile-Net on ImageNet compared to baselines.

报告人信息

Dr. Shi Gu obtained his Bachelor’s degree in Mathematics from Tsinghua University in 2011 and PhD degree in Applied Mathematics and Computational Science from University of Pennsylvania in 2016. In 2017, he joined University of Electronic Science and Technology of China as a professor in the Department of Computer Science through the national young talent program. His research interests integrate neuroscience with machine learning. Specially he is interested in developing novel computational methods to explain the cognition and building cognition-inspired algorithms for efficient neural networks. His works have been published on Nature Communications, PNAS, TMI, ICLR, etc.