BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//7.2.3.1//EN
BEGIN:VEVENT
UID:756@lincs.fr
DTSTART;TZID=Europe/Paris:20230412T150000
DTEND;TZID=Europe/Paris:20230412T160000
DTSTAMP:20230413T090818Z
URL:https://www.lincs.fr/events/distributed-function-computation-over-netw
 orks/
SUMMARY:Distributed Function Computation over Networks
DESCRIPTION:Large-scale distributed computing systems\, such as MapReduce\,
 Spark\, or distributed deep networks\, are critical for parallelizing the
 execution of computational tasks. Nevertheless\, a struggle between
 computation and communication complexity lies at the heart of distributed
 computing. There has been recently a substantial effort to address this
 problem for a class of functions\, such as distributed matrix
 multiplication\, distributed gradient coding\, linearly separable
 functions\, etc. The optimal cost has been achieved under some
 constraints\, based mainly on ideas of linear separability of the tasks and
 linear space intersections. Motivated by the same challenge\, we propose a
 novel distributed computing framework where a master seeks to compute an
 arbitrary function of distributed datasets in an asymptotically lossless
 manner. Our approach exploits the notion of characteristic graphs\, which
 have been widely utilized by Shannon\, K¨orner\, and Witsenhausen to
 derive the rate lower bounds for computation\, and later by Alon-
 Orlitsky\, Orlitsky-Roche\, Doshi-Shah-M´edard\, and Feizi-M´edard\, to
 resolve some well known distributed coding and communication problems\,
 allowing for lowered communication complexity and even for a) correlated
 data\, b) a broad class of functions\, and c) well-known topologies. The
 novelty of our approach lies in accurately capturing the
 communication-computation cost tradeoff by melding the notions of
 characteristic graphs and distributed placement\, to provide a natural
 generalization of distributed linear function computation\, thus elevating
 distributed gradient coding and distributed linear transform to the realm
 of distributed computing of any function. In toy scenarios\, we demonstrate
 gains up to %70 over fully distributed solutions and an approximation ratio
 of 2 within the optimal centralized rate.
CATEGORIES:Seminars,Youtube
LOCATION:Room 4B01\, 19 place Marguerite Perey\, Palaiseau\, France
X-APPLE-STRUCTURED-LOCATION;VALUE=URI;X-ADDRESS=19 place Marguerite Perey\,
 Palaiseau\, France;X-APPLE-RADIUS=100;X-TITLE=Room 4B01:geo:0,0
END:VEVENT
BEGIN:VTIMEZONE
TZID:Europe/Paris
X-LIC-LOCATION:Europe/Paris
BEGIN:DAYLIGHT
DTSTART:20230326T030000
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
END:DAYLIGHT
END:VTIMEZONE
END:VCALENDAR