Stanford Seminar - Song Han of Stanford University



2
7324

"Deep Compression and EIE: Deep Neural Network Model Compression and Hardware Acceleration" - Song Han of Stanford University Support for the Stanford Colloquium on Computer Systems Seminar Series provided by the Stanford Computer Forum. Speaker Abstract and Bio can be found here: http://ee380.stanford.edu/Abstracts/160106.html For more info about the EE380 series: http://ee380.stanford.edu Colloquium on Computer Systems Seminar Series (EE380) presents the current research in design, implementation, analysis, and use of computer systems. Topics range from integrated circuits to operating systems and programming languages. It is free and open to the public, with new lectures each week. Learn more: http://bit.ly/WinYX5 0:00 Introduction 1:07 This Talk 1:33 Deep Learning: Next Wave of Al 2:26 Applications 6:51 Solver 1: Deep Compression 7:45 Solve 2: EIE Accelerator 9:08 Compression Pipeline: Overview 10:04 1. Pruning 10:33 Pruning: Motivation 11:32 Pruning: Result on 4 Covnets 12:10 Pruning: AlexNet 14:04 AlexNet & VGGNet 15:11 Mask Visualization 20:47 Speedup (FC layer) 23:03 Energy Efficiency (FC layer) 23:26 2. Quantization and Weight Sharing 24:03 Weight Sharing: Overview 25:44 Finetune Centroids 27:17 Quantization: Result 28:52 Pruning and Quantization Works Well Together 30:49 Huffman Coding 31:45 Deep Compression Result on 4 Convnets 33:19 AlexNet: Breakdown 33:50 Comparison with other Compression Methods 34:06 Conclusion 35:00 A Model Compression Tool for Developers 36:08 ASIC Accelerator that Runs DNN on Mobile 38:58 Distribute Storage and Processing 39:52 Inside each PE 40:56 Evaluation 42:53 Baseline and Benchmark 43:17 Layout of an EIE PE 44:26 Result: Speedup / Energy Efficiency 47:30 Scalability 48:11 Useful Computation / Load Balance 49:53 Design Space Exploration 51:59 Hardware for Deep Learning

Published by: Stanford Online Published at: 8 years ago Category: آموزشی