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:20260402T024532Z
LOCATION:Level 2 Lobby
DTSTART;TZID=America/Los_Angeles:20250623T180000
DTEND;TZID=America/Los_Angeles:20250623T190000
UID:dac_DAC 2025_sess262_RESEARCH2337@linklings.com
SUMMARY:Efficient Edge AI Learning with Equilibrium Propagation: A Practic
 al Solution For Gradient Computation
DESCRIPTION:Mohamed Watfa (LIRMM / CNRS), Alberto Garcia-Ortiz (ITEM (U.Br
 emen)), and Gilles Sassatelli (LIRMM / CNRS)\n\nThe rapid growth of smart 
 devices and sensors has led to an overwhelming\nincrease in data generatio
 n, pushing current network infrastructure to its\nlimits and threatening t
 he scalability of cloud-based processing. Edge machine\nlearning, which pr
 ocesses data locally on devices, presents a viable solution\nto reduce net
 work load and latency. However, deploying deep learning at the\nedge remai
 ns difficult due to the limited memory and computational capacity of\nthes
 e devices which mostly precludes on-device/on-site training. Equilibrium\n
 propagation (EP) has emerged as a promising alternative to backpropagation
 ,\nleveraging analog processing and device physics for energy-efficient le
 arning.\nYet, its practical implementation is hindered by challenges such 
 as voltage\nvariations and the need for energy-efficient circuits capable 
 of gradient\ncomputation at a sufficient level of accuracy. Existing solut
 ions rely on\nimpractical idealized models. In this work, we introduce a n
 ovel method to\naddress the problem of the wide dynamic range of the volta
 ge variation to avoid\nthe use of expensive low-noise amplifiers, and prop
 ose an innovative\ntransistor-level switched-capacitor circuit to compute 
 gradients in accordance\nwith the EP rule. Additionally, our design suppor
 ts batching, a key requirement\nfor stable training that is often overlook
 ed. We validate our approach on the MNIST\ndataset, demonstrating a practi
 cal, energy-efficient EP circuit that operates\nwithin real hardware const
 raints.\n\nTracks: DES5: Emerging Device and Interconnect Technologies\n\n
END:VEVENT
END:VCALENDAR