BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Denver
X-LIC-LOCATION:America/Denver
BEGIN:DAYLIGHT
TZOFFSETFROM:-0700
TZOFFSETTO:-0600
TZNAME:MDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0600
TZOFFSETTO:-0700
TZNAME:MST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20200129T163603Z
LOCATION:502-503-504
DTSTART;TZID=America/Denver:20191117T114500
DTEND;TZID=America/Denver:20191117T121000
UID:submissions.supercomputing.org_SC19_sess101_ws_dls103@linklings.com
SUMMARY:DC-S3GD: Delay-Compensated Stale-Synchronous SGD for Large-Scale D
 ecentralized Neural Network Training
DESCRIPTION:Workshop\n\nDC-S3GD: Delay-Compensated Stale-Synchronous SGD f
 or Large-Scale Decentralized Neural Network Training\n\nRigazzi\n\nData pa
 rallelism has become the de facto standard for training Deep Neural Networ
 k on mul- tiple processing units. In this work we propose DC- S3GD, a dece
 ntralized (without Parameter Server) stale-synchronous version of the Dela
 y-Compensated Asynchronous Stochastic Gradient Descent (DC- ASGD) algorith
 m. In our approach, we allow for the overlap of computation and communicat
 ion, by averaging in parameter space and compensating the inherent error w
 ith a first-order correction of the locally computed gradients. We prove t
 he effectiveness of our approach by training Convolutional Neural Network 
 with large batches and achieving state-of- the-art results.\n\nTag: Worksh
 op Reg Pass, Deep Learning, Scientific Computing\n\nRegistration Category:
  Workshop Reg Pass, Deep Learning, Scientific Computing
URL:https://sc19.supercomputing.org/presentation/?id=ws_dls103&sess=sess10
 1
END:VEVENT
END:VCALENDAR