BEGIN:VCALENDAR
VERSION:2.0
PRODID:icalendar-ruby
CALSCALE:GREGORIAN
X-WR-CALNAME:MIT Probability Seminar
X-WR-TIMEZONE:Eastern Time (US & Canada)
BEGIN:VEVENT
DTSTAMP:20260712T021842Z
UID:tag:localist.com\,2008:EventInstance_39138182817133
DTSTART:20220307T211500Z
DTEND:20220307T221500Z
DESCRIPTION:Speaker: Mark Sellke (Stanford University)\n\nTitle: Algorithmi
 c Thresholds in Mean-Field Spin Glasses\n\nAbstract: I will explain recent
  progress on computing approximate ground states of mean-field spin glass 
 Hamiltonians\, which are certain random functions in high dimension. While
  the asymptotic ground state energy OPT is given by the famous Parisi form
 ula\, the landscape of these functions often include many bad local optima
  which impede optimization by efficient algorithms. In the positive direct
 ion\, I will explain algorithms based on approximate message passing which
  asymptotically achieve a value ALG given by an extended Parisi formula. T
 he case ALG=OPT has a "no overlap gap" or "full replica symmetry breaking"
  interpretation\, but in general these algorithms may fail to reach asympt
 otic optimality. In the negative direction\, I will explain why no algorit
 hm with suitably Lipschitz dependence on the random disorder can surpass t
 he threshold ALG. This result applies to many standard optimization algori
 thms\, such as gradient descent and its variants on dimension-free time sc
 ales. Based on joint works with Ahmed El Alaoui\, Brice Huang\, and Andrea
  Montanari.
GEO:42.358262;-71.090045
LOCATION:Building 2\, 2-147
SUMMARY:MIT Probability Seminar
URL;VALUE=URI:https://calendar.mit.edu/event/mit_probability_seminar_202203
 07
CATEGORIES:Conferences/Seminars/Lectures
END:VEVENT
END:VCALENDAR