BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//UM//UM*Events//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:America/Detroit
TZURL:http://tzurl.org/zoneinfo/America/Detroit
X-LIC-LOCATION:America/Detroit
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:20070311T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:20071104T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20241008T111956
DTSTART;TZID=America/Detroit:20241018T100000
DTEND;TZID=America/Detroit:20241018T110000
SUMMARY:Workshop / Seminar:Statistics Department Seminar Series: George Michailidis\, Professor\, Department of Statistics and Data Science\, UCLA
DESCRIPTION:Abstract: Bilevel optimization (BLO) is an optimization problem that involves two levels of hierarchy (i.e.\, upper and lower levels)\, wherein obtaining the solution to the upper-level problem requires solving the lower-level one. BLO has recently attracted a lot of attention due to novel applications in machine learning that involve optimizing nested objective functions\, including hyperparameter tuning and related automated ML tasks\, few shot robust\, representation\, coreset and reinforcement learning. In the first part of the talk\, a very brief overview of the BLO problem\, selected machine learning applications that can be cast as BLO problems and algorithms to obtain its solution will be given.\n\nIn the second part of the talk\, we introduce the decentralized version of the BLO problem (without assuming a star-shape network topology) and present a penalty function-based decentralized algorithm with theoretical guarantees under both convex and non-convex assumptions for the upper level functions. Specifically\, a distributed alternating gradient-type algorithm for solving consensus BLO over a decentralized network is developed. A key feature of the proposed algorithm is the estimation of the hyper-gradient of the penalty function through decentralized computation of matrix-vector products and few vector communications. Further\, the theoretical results highlight improvements in the iteration complexity of decentralized BLO. Empirical results on selected ML problems demonstrate that the proposed method performs well in real-world settings.\n\nhttps://georgemichailidis.github.io/
UID:124540-21853169@events.umich.edu
URL:https://events.umich.edu/event/124540
CLASS:PUBLIC
STATUS:CONFIRMED
CATEGORIES:seminar
LOCATION:West Hall - 340
CONTACT:
END:VEVENT
END:VCALENDAR