모바일 메뉴 닫기
 

Research

Research Areas & Labs

Title
Seminar [05/03] On-Device AI Systems: From DNN Model Compressions to Machine Learning Accelerators
Date
2019.04.29
Writer
전기전자공학부
게시글 내용

< BK21+ BEST Seminar Series Announcement> 


Time and Date : 16:00 ~ 17:00 Friday 05/03/2019

Place : D403, Engineering Building #4

Title : On-Device AI Systems: From DNN Model Compressions to Machine Learning Accelerators
Abstract:
AI is one of the hottest keywords recently in computer societies. Its great successes in computer vision and speech recognition have encouraged many researchers and engineers to apply AI to almost every area of science and engineering. Though key AI algorithms were originally developed in decades ago, it was big data availability and hardware acceleration that really made AI so successful today.

With On-Device AI systems, we aim to deliver transparent AI user experience on user devices without connecting to cloud servers. To enable AI functionalities on mobile and consumer devices, a system-level holistic optimization from algorithm to chip is crucial to overcome AI application’s compute, memory and power requirements. We have to develop small but accurate DNN models, hardware accelerators to run the models, and runtime and compiler to manage the accelerators efficiently.

First, I will brief discuss background knowledge to understand DNN accelerators. I would like to cover important academic achievements. And, I like to show the landscape of NPU industry, introducing various NPU chips and explaining their pros and cons. NPU technology makes progress so fast and the gap between academia and industry is small. Many commercial NPUs such as Huawei’s Kirin NPU were initially made by university researchers. Then, I will introduce our efforts for DNN Model compression and DNN compiler and runtime acceleration. I would like to show model compression techniques such as pruning, quantization, and matrix factorization. And, I like to discuss how to efficiently exploit computing resources on a mobile SoC via a runtime technique. Finally, I will conclude the talk by discussing future works. Throughout the talk, I would like to provide the audience chances to grasp key ideas on on-device AI systems.


 

Presenter: Daehyun Kim, Samsung Research

Host: Prof. Ro, Wonwoo, Yonsei EEE