TUESDAY June 26, 1:30pm - 3:00pm | Room 3014
TOPIC AREA: MACHINE LEARNING/AI, DESIGN
KEYWORD: ARCHITECTURE & SYSTEM DESIGN, LOW-POWER & RELIABILITY
EVENT TYPE: RESEARCH REVIEWED

SESSION 19
Watch Your Bits: Precision and Fault Tolerance in Deep Learning
Chair:
Muhammad Shafique - Technische Univ. Wien, Vienna, Austria
Co-Chair:
Ahmed Hemani - KTH Royal Institute of Technology, Kista, Sweden
Designers must account for the effect of error and imprecision on DNN behavior, especially since these characteristics of DNN can be leveraged to improve performance and energy. Ares presents a fault-injection framework for estimating the resilience of DNNs to permanent hardware faults. DeepN-JPEG revisits JPEG quantization in order to improve classification accuracy when using compressed images. ThUnderVolt enables voltage underscaling of DNN accelerators by tolerating timing errors. Loom presents an accelerator that exploits the variable precision required by different layers of a CNN, increasing performance by reducing precision.

19.1*Ares: A Framework for Quantifying the Resilience of Deep Neural Networks
 Speaker: Brandon Reagen - Harvard Univ., Cambridge, MA
 Authors: Brandon Reagen - Harvard Univ., Cambridge, MA
Udit Gupta - Harvard Univ., Cambridge, MA
Lillian Pentecost - Harvard Univ., Cambridge, MA
Paul N. Whatmough - Arm, Ltd. & Harvard Univ., Cambridge, MA
Sae Kyu Lee - IBM T.J. Watson Research Center & Harvard Univ., Cambridge, MA
Niamh Mulholland - Harvard Univ., Cambridge, MA
Gu-Yeon Wei - Harvard Univ., Cambridge, MA
David Brooks - Harvard Univ., Cambridge, MA
19.2DeepN-JPEG: A Deep Neural Network Favorable JPEG-based Image Compression Framework
 Speaker: Zihao Liu - Florida International Univ., Miami, FL
 Authors: Zihao Liu - Florida International Univ., Miami, FL
Tao Liu - Florida International Univ., Miami, FL
Wujie Wen - Florida International Univ., Miami, FL
Lei Jiang - Indiana Univ., Bloomington, IN
Jie Xu - University of Miami, FL
Yanzhi Wang - Syracuse Univ., Syracuse, NY
Gang Quan - Florida International Univ., Miami, FL
19.3ThUnderVolt: Enabling Aggressive Voltage Underscaling and Timing Error Resilience for Energy Efficient Deep Learning Accelerators
 Speaker: Jeff Zhang - New York Univ., Brooklyn, NY
 Authors: Jeff Zhang - New York Univ., Brooklyn, NY
Kartheek Rangineni - Indian Institute of Technology Kanpur, India
Zahra Ghodsi - New York Univ., Brooklyn, NY
Siddharth Garg - New York Univ., New York, NY
19.4Loom: Exploiting Weight and Activation Precisions to Accelerate Convolutional Neural Networks
 Speaker: Alberto Delmas Lascorz - Univ. of Toronto, ON, Canada
 Authors: Sayeh Sharifymoghaddam - Univ. of Toronto, ON, Canada
Alberto Delmas Lascorz - Univ. of Toronto, ON, Canada
Patrick Judd - Univ. of Toronto, ON, Canada
Andreas Moshovos - Univ. of Toronto, ON, Canada


* Indicates Best Paper Candidate