A New Path Towards Efficient, Sustainable and Secure Deep Learning System Design

Abstract

Thanks to recent machine learning model innovation and computing hardware advancement, deep neural network or deep learning, is nowadays becoming the technique foundation of many real-world applications, such as computer vision, robotics, autonomous vehicles, medical imaging and diagnostics etc. However, the real deployment of deep learning, as one of the most active areas in computing engineering, still faces critical challenges like low performance bottlenecked by data offloading, instable inference quality of energy-efficient emerging process-in-memory (PIM) based accelerators because of device limitations and ever-increasing security concern due to adversarial data with imperceptible perturbations. In this talk, I will address these issues from several radically different perspectives on computing and security. The talk starts with the development of a “machine vision” (NOT “human vision”) guided image compression framework tailored for high-performance deep learning service with guaranteed accuracy for the first time, based on an insightful understanding about the difference between deep learning (or “machine vision”) and human vision on image perception. Then I will discuss error correction code and transfer learning inspired neural network classifier design to mitigate the stability issue of PIM based emerging accelerators built upon memristor device without involving expensive chip-specific defect map calibration or training-from-scratch. I will also show how to seamlessly integrate the defense into the low cost data compression in a rigorous manner to achieve by far the best defense efficiency (among the input-based countermeasures) against emerging adversarial-example attacks, with almost no accuracy drop of benign data. Our prospects on the research of deep learning, e.g. our advocated deep learning security from a system perspective, will be also given at the end of this talk, offering the audiences an alternative thinking about developing more efficient, sustainable and secure deep learning systems.

Biography

Dr. Wujie Wen is currently an assistant professor in the department of Electrical and Computer Engineering at Florida International University (FIU), Miami, FL. He received his Ph.D. from University of Pittsburgh in 2015. He earned his B.S. and M.S. degrees in electronic engineering from Beijing Jiaotong University and Tsinghua University, Beijing, China, in 2006 and 2010, respectively. Before he joined FIU, he also worked with AMD and Broadcom for various engineer and intern positions. His current research interests include deep learning hardware acceleration/security, neuromorphic computing, and circuit-architecture design for emerging memory technologies. His works have been published in top-tie conferences and journals (e.g. HPCA, DAC, ICCAD, DATE, ICPP, HOST, ECCV, AAAI, TC, TCAD etc). Dr. Wen is the associate editor of Neurocomputing and serves as the General Chair of ISVLSI 2019 (Miami), Technical Program Chair of ISVLSI 2018 (Hong Kong), as well as program committee for many conferences such as DAC, ICCAD, ASP-DAC etc. He received best paper nominations from ASP-DAC2018, ICCAD2018, DATE2016 and DAC2014. He was also the recipient of the 49th DAC A. Richard Newton Graduate Scholarship, the most prestigious Ph.D. scholarship (one awardee per year) in EDA society and 2015 DAC Ph.D. forum best poster presentation. His research is sponsored by NSF, AFRL and Florida Center for Cybersecurity etc.