BEGIN:VCALENDAR
VERSION:2.0
CALSCALE:GREGORIAN
PRODID:UW-Madison-Physics-Events
BEGIN:VEVENT
SEQUENCE:4
UID:UW-Physics-Event-5934
DTSTART:20200909T160000Z
DTEND:20200909T173000Z
DTSTAMP:20260415T193202Z
LAST-MODIFIED:20200905T011400Z
LOCATION:Please register for this online event: http://physicsmeetsml.
 org
SUMMARY:Insights on gradient-based algorithms in high-dimensional lear
 ning\, Physics ∩ ML Seminar\, Lenka Zdeborova\, Université Paris-Sa
 clay
DESCRIPTION:Gradient descent algorithms and their noisy variants\, suc
 h as the Langevin dynamics or multi-pass SGD\, are at the center of at
 tention in machine learning. Yet their behaviour remains perplexing\, 
 in particular in the high-dimensional non-convex setting. In this talk
 \, I will present several high-dimensional and (mostly) non-convex sta
 tistical learning problems in which the performance of gradient-based 
 algorithms can be analysed down to a constant. The common point of the
 se settings is that the data come from a probabilistic generative mode
 l leading to problems for which\, in the high-dimensional limit\, stat
 istical physics provides exact closed solutions for the performance of
  the gradient-based algorithms. The covered settings include the spike
 d mixed matrix-tensor model\, the perceptron or phase retrieval.
URL:https://www.physics.wisc.edu/events/?id=5934
END:VEVENT
END:VCALENDAR