BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Computer-aided proofs in first-order optimization\, with applicati
 ons to error feedback
DTSTART:20260529T151500
DTEND:20260529T161500
DTSTAMP:20260526T170119Z
UID:662ec8f5d23060a3da78da11f4efc8de0b1205ebdbb07dfe9e1470e8
CATEGORIES:Conferences - Seminars
DESCRIPTION:Aymeric Dieuleveut\, Ecole Polytechnique\, Institut Polytechn
 ique de Paris\nFirst-order methods are widely used in optimization and mac
 hine learning\, and their behavior is often analyzed through the spectrum 
 of worst case convergence rates. Obtaining such guarantees is often diffic
 ult and both time consuming and error-prone. Starting with the work of Dro
 ri and Teboulle (2014)\, novel techniques have been used to gain numerical
  insights\, leading to the release of various performance estimation (PE) 
 software. \n\nIn this talk\, I will show how various computer-aided techn
 iques can be used to study first-order optimization methods in a systemati
 c way. From performance estimation problems with automated Lyapunov discov
 ery\, to symbolic regression and computer algebra systems\, novel tools co
 mpletely reshape the way we approach theory of optimization. \nAs a main 
 example\, I will focus on error feedback methods used with compressed comm
 unication in distributed optimization. While error feedback has been widel
 y studied\, existing theory often provides untight (thus unreliable)  bou
 nds. I will present tight analyses with matching lower bounds that allow a
  fair comparison between error feedback schemes and standard compressed gr
 adient descent\, and help explain when error feedback is useful and when i
 t is not.\n\nOverall\, the talk aims to show how various computer-aided pr
 oofs can lead to clearer and more reliable insights into first-order optim
 ization methods.\n\nBased on :\nA Tight Theory of Error Feedback Algorithm
 s in Distributed Optimization\, DB Thomsen\, A Taylor\, A Dieuleveut\nInt
 ernational Conference on Machine Learning (ICML 2026)\nTight analyses of f
 irst-order methods with error feedback\, DB Thomsen\, A Taylor\, A Dieulev
 eut\nAdvances in Neural Information Processing Systems (NeurIPS) (2025).\n
 PEPit: computer-assisted worst-case analyses of first-order optimization m
 ethods in Python\nB Goujaud\, C Moucer\, F Glineur\, JM Hendrickx\, AB Tay
 lor\, A Dieuleveut\nMathematical Programming Computation 16 (3)\, 337-367
   (2024)\n(eventually - as a detour) Provable non-accelerations of the h
 eavy-ball method B Goujaud\, A Taylor\, A Dieuleveut\, Mathematical Prog
 ramming\, 1-59  (2025)\n 
LOCATION:CM 1 517 https://plan.epfl.ch/?room==CM%201%20517
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR