BEGIN:VCALENDAR
VERSION:2.0
PRODID:researchseminars.org
CALSCALE:GREGORIAN
X-WR-CALNAME:researchseminars.org
BEGIN:VEVENT
SUMMARY:Ben Adcock (Simon Fraser University)
DTSTART:20230126T233000Z
DTEND:20230127T003000Z
DTSTAMP:20260513T201629Z
UID:SFUOR/8
DESCRIPTION:Title: <a href="https://researchseminars.org/talk/SFUOR/8/">Re
 starts Subject to Approximate Sharpness: a Parameter-free and Optimal Sche
 me for Accelerating First-order Methods</a>\nby Ben Adcock (Simon Fraser U
 niversity) as part of PIMS-CORDS SFU Operations Research Seminar\n\nLectur
 e held in ASB 10908.\n\nAbstract\nSharpness is a generic assumption in con
 tinuous optimization that bounds the distance to the set of minimizers in
  terms of the suboptimality in the objective function. It leads to the ac
 celeration of first-order optimization methods via so-called restarts. How
 ever\, sharpness involves problem-specific constants that are typically u
 nknown\, and previous restart schemes often result in reduced convergence
  rates. Such schemes are also challenging to apply in the presence of nois
 e or approximate model classes (e.g.\, in compressed sensing or machine l
 earning problems). In this talk\, we introduce the notion of approximate 
 sharpness\, a generalization of sharpness that incorporates an unknown co
 nstant perturbation to the objective function error. By employing a new ty
 pe of search over the unknown constants\, we then describe a restart sche
 me that applies to general first-order methods. Our scheme maintains the 
 same convergence rate as when assuming knowledge of the constants. Moreov
 er\, for broad classes of problems\, it gives rates of convergence which e
 ither match known optimal rates or improve on previously established rate
 s. Finally\, we demonstrate the practical efficacy of this scheme on appl
 ications including sparse recovery\, compressive imaging and feature sele
 ction in machine learning.\n\nThis is joint work with Matthew J. Colbrook 
 (Cambridge) and Maksym Neyra-Nesterenko (SFU). The corresponding paper ca
 n be found here: https://arxiv.org/abs/2301.02268\n
LOCATION:https://researchseminars.org/talk/SFUOR/8/
END:VEVENT
END:VCALENDAR
