报告题目：RevisitFast Algorithms for Convolutional Neural Networks

报告人：Dr. Zhen Jia，AWS AI, Amazon

报告时间：2019年12月16日下午13:30

报告地点：中心校区物理楼345讲学厅

报告摘要：Asconvolutional neural network (ConvNet) regained its popularity, several fastapproaches were proposed to speed up the computation of its most time-consumingconvolutional layers. Recently, it has been demonstrated that Winograd basedconvolution can greatly reduce the computation complexity of a convolution. Acommon belief is that Winograd based algorithms should be faster than FFT-basedones, which were the most frequently used methods to reduce the computationbefore Winograd proposed.

Inthis talk, I will first present our work on optimizing the Winograd basedconvolution. Our implementation achieves high hardware utilization through aseries of optimizations and runs on average more than 3x faster than otherstate-of-the-art CPU implementations on the benchmarked layers. Then, I willrevisit FFT. We propose a performance model based on the idea of Rooflinemodel to analyze and compare the performance of FFT and Winograd baseconvolutions on modern processors. To validate the predictions of our model, wealso implement FFT based convolutions with the same optimization methods usedin Winograd. The results of our experiments verify our model's prediction anddemonstrate that Winograd is not always faster than FFT convolution.

报告人简介：ZhenJia is an applied scientist in the AWS AI team of Amazon, CA, USA. Beforejoining Amazon. He was a Postdoctoral Research Associate at PrincetonUniversity in Computer Science Department. His research interests includeconvolutional neural network accelerating, workload characterization,benchmarking, and performance optimization at system level. He received thePh.D degree from Institute of Computing Technology (ICT), Chinese Academy ofSciences (CAS). He has published over 30 papers in top computer systemconferences and journals like HPCA, PPoPP, ICS, PACT, TPDS, etc. Hispublications have been cited 1,100+ times to date.

主办单位：尊龙凯时物理学院

中国声学学会物理声学分会、检测声学分会

高等学校计算物理研究会

尊龙凯时省计算物理学会