BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
X-LIC-LOCATION:America/Los_Angeles
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20211208T164410Z
LOCATION:3014
DTSTART;TZID=America/Los_Angeles:20211208T113000
DTEND;TZID=America/Los_Angeles:20211208T115000
UID:dac_DAC 2021_sess156_RESEARCH1126@linklings.com
SUMMARY:Exact Neural Networks from Inexact Multipliers via Fibonacci Weigh
 t Encoding
DESCRIPTION:Research Manuscript\n\nExact Neural Networks from Inexact Mult
 ipliers via Fibonacci Weight Encoding\n\nSimon\n\nEdge devices must suppor
 t computationally demanding algorithms, such as neural networks, within ti
 ght area/energy budgets. While approximate computing is demonstrated to al
 leviate these constraints, limiting induced errors remains an open challen
 ge. In this paper, we propose a hardware/software co-design solution via a
 n inexact multiplier, reducing area/power-delay-product requirements by 73
 /43%, respectively, while still computing exact results when one input is 
 a Fibonacci encoded value. We then introduce a retraining strategy to quan
 tize neural network weights to Fibonacci encoded values, ensuring exact co
 mputation during inference. We benchmark our strategy on Squeezenet 1.0, D
 enseNet-121, and ResNet-18, measuring accuracy degradations of only 0.4/1.
 1/1.7%, respectively.\n\nTopic: Design\n\nKeyword: Approximate Computing f
 or AI/ML\n\nVirtual Program: Presented In-Person
END:VEVENT
END:VCALENDAR