BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Scalable constrained optimization
DTSTART:20190826T160000
DTEND:20190826T180000
DTSTAMP:20260429T094700Z
UID:e70e9ea874cbf22c5d2e0dddc6365bb4639c58a20154c9334d224c9b
CATEGORIES:Conferences - Seminars
DESCRIPTION:Maria-Luiza Vladarean\nEDIC candidacy exam\nExam president: Pr
 of. Ali H. Sayed\nThesis advisor: Prof. Volkan Cevher\nCo-examiner: Prof. 
 Martin Jaggi\n\nAbstract\nThe recent surge of data availability poses a si
 gnificant burden on optimization algorithms. Methods are required to be ro
 bust in the presence of noise and scalable\, i.e. provably fast and space-
 efficient. In the particular case of constrained optimization\, the latter
  property comes in the form of projection-free methods\, stochastic constr
 aints\, or distributable schemes.\n\nThis proposal discusses three concept
 ually different approaches to constrained minimization problems. We first 
 present a unified convergence analysis for classical Augmented Lagrangian 
 schemes. We then study a method that leverages a probabilistic framework t
 o satisfy linear inclusion constraints - possibly infinitely many - almost
  surely. Finally\, we discuss the zeroth-order optimization perspective on
  constrained problems and the scalability challenges faced in the absence 
 of gradient information. We conclude this proposal by outlining some preli
 minary results obtained during our study of almost-sure constraints under 
 the conditional gradient framework.\n\nBackground papers\nAlmost surely co
 nstrained convex optimization\, by Olivier Fercoq\, Ahmet Alacaoglu\, Ion
  Necoara\, Volkan Cevher.\nLagrangian methods for composite optimizati
 on\, by Shoham Sabach\, Marc Teboulle.\nZeroth-order Nonconvex Stochastic 
 Optimization: Handling Constraints\, High-Dimensionality and Saddle-Points
 \, by Krishnakumar Balasubramanian and Saeed Ghadimi.\n \n 
LOCATION:DIA 004 https://plan.epfl.ch/?room==DIA%20004
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR