BEGIN:VCALENDAR
VERSION:2.0
PRODID:researchseminars.org
CALSCALE:GREGORIAN
X-WR-CALNAME:researchseminars.org
BEGIN:VEVENT
SUMMARY:Asu Ozdaglar (Massachusetts Institute of Technology)
DTSTART:20201116T143000Z
DTEND:20201116T153000Z
DTSTAMP:20260423T021012Z
UID:OWOS/30
DESCRIPTION:Title: <a href="https://researchseminars.org/talk/OWOS/30/">Ro
 bustness in Machine Learning and Optimization: A Minmax Approach</a>\nby A
 su Ozdaglar (Massachusetts Institute of Technology) as part of One World O
 ptimization seminar\n\n\nAbstract\nMinmax problems arise in a large number
  of problems in optimization\, including worst-case design\, duality theor
 y\, and zero-sum games\, but also have become popular in machine learning 
 in the context of adversarial robustness and Generative Adversarial Networ
 ks (GANs). This talk will review our recent work on solving minmax problem
 s using discrete-time gradient based optimization algorithms. We focus on 
 Optimistic Gradient Descent Ascent (OGDA) and Extra-gradient (EG) methods\
 , which have attracted much attention in the recent literature because of 
 their superior empirical performance in GAN training.  We show that OGDA a
 nd EG can be seen as approximations of the classical proximal point method
  and use this interpretation to establish convergence rate guarantees for 
 these algorithms. These guarantees are provided for the ergodic (averaged)
  iterates of the algorithms. We also consider the last iterate of EG  and 
 present convergence rate guarantees for the last iterate for smooth convex
 -concave saddle point problems. We finally turn to analysis of generalizat
 ion properties of gradient based minmax algorithms using the algorithmic s
 tability framework defined by Bousquet and Elisseeff. Our generalization a
 nalysis suggests superiority of gradient descent ascent (GDA) compared to 
 GDmax algorithm (which involves exact solution of the maximization problem
  at each iteration) in the nonconvex-concave case provided that similar le
 arning rates are used in the descent and ascent steps.\n\nThe address and 
 password of the zoom room of the seminar are sent by e-mail on the mailing
 list of the seminar one day before each talk\n
LOCATION:https://researchseminars.org/talk/OWOS/30/
END:VEVENT
END:VCALENDAR