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:20240626T180033Z
LOCATION:3003\, 3rd Floor
DTSTART;TZID=America/Los_Angeles:20240626T134500
DTEND;TZID=America/Los_Angeles:20240626T140000
UID:dac_DAC 2024_sess125_RESEARCH152@linklings.com
SUMMARY:Deep Reorganization: Retaining Residuals in TinyML
DESCRIPTION:Research Manuscript\n\nHashan Mendis and Chih-Kai Kang (Academ
 ia Sinica), Chun-Han Lin (National Taiwan Normal University), Ming-Syan Ch
 en (National Taiwan University), and Pi-Cheng Hsiu (Academia Sinica)\n\nDe
 signing intelligent, tiny devices with limited memory is immensely challen
 ging, exacerbated by the additional memory requirement of residual connect
 ions in deep neural networks. In contrast to existing approaches that elim
 inate residuals to reduce peak memory usage at the cost of significant acc
 uracy degradation, this paper presents DERO, which reorganizes residual co
 nnections by leveraging insights into the types and interdependencies of o
 perations across residual connections. Evaluations were conducted across d
 iverse model architectures designed for common computer vision application
 s. DERO consistently achieves peak memory usage comparable to plain-style 
 models without residuals, while maintaining the accuracy of the original m
 odels with residuals.\n\nTopic: AI, Design\n\nKeyword: AI/ML System and Pl
 atform Design\n\nSession Chairs: Amin Firoozshahian (Rain AI) and Thierry 
 Tambe (Stanford University)
END:VEVENT
END:VCALENDAR