BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//7.4.0.1//EN
TZID:America/New_York
X-WR-TIMEZONE:America/New_York
BEGIN:VEVENT
UID:938@fds.yale.edu
DTSTART;TZID=America/New_York:20260227T130000
DTEND;TZID=America/New_York:20260227T140000
DTSTAMP:20260203T173138Z
URL:https://fds.yale.edu/events/fds-seminar-ilyas-fatkhullin-ethz/
SUMMARY:FDS Seminar: Ilyas Fatkhullin (ETHZ)\, "Toward Dependable Reinforce
 ment Learning: Hidden Structures and Safe Optimization"
DESCRIPTION:\nZoom Password: 123\n\n\n\nAbstract: Modern reinforcement lear
 ning (RL) has enabled impressive progress in sequential decision-making\, 
 yet state-of-the-art performance often requires enormous engineering effor
 ts and still lacks reliable\, theory-backed explanations. This talk develo
 ps an optimization-centric perspective on RL aimed at reliability: algorit
 hms that remain stable in high-dimensional\, stochastic\, and inherently n
 on-convex settings. I will walk you through three important pillars: (1) n
 on-convex landscapes and hidden structures\, (2) robustness and data-effic
 iency\, and (3) scalable distributed learning. Going deeper into the first
  pillar\, I show how RL objectives that appear intractable can nevertheles
 s exhibit exploitable geometry — settings where non-convexity hides an i
 mplicit convex backbone. Such hidden convex structures yield global conver
 gence guarantees for simple stochastic optimization procedures closely con
 nected to policy-gradient-style methods. Building on this foundation\, I a
 ddress a critical challenge of safety in RL — respecting intrinsically n
 on-convex constraints. I outline a globally convergent method based on a s
 hifted inexact proximal-point framework that can drive iterates toward fea
 sibility and optimality without relying on smoothness or constrained quali
 fication assumptions\, and demonstrate its practical behavior in safe expl
 oration.I conclude by outlining how (2) robustness and (3) scalability con
 nect these geometric insights to reliable learning systems deployed under 
 uncertainty and resource constraints\, and discuss future directions.\n\n\
 n\nSpeaker Bio: Ilyas Fatkhullin is a Ph.D. candidate in Computer Science 
 at ETH Zurich and an ETH AI Center Fellow. His research lies at the inters
 ection of optimization theory and reinforcement learning\, developing prin
 cipled methods for reliable learning in high-dimensional\, non-convex\, an
 d stochastic settings\, including hidden convexity\, heavy-tailed noise\, 
 and distributed training. His work has appeared in venues such as NeurIPS\
 , ICML\, and AISTATS\, and in journals including SIAM Journal on Optimizat
 ion\, SIAM Journal on Control and Optimization\, and JMLR. Special highlig
 hts include an oral presentation at NeurIPS 2021 (main track)\, a spotligh
 t presentation at ICML 2022 (main track)\, and oral presentations at the N
 eurIPS 2025 OPT and COML workshops.\n
CATEGORIES:FDS Events,Postdoctoral Applicants
END:VEVENT
BEGIN:VTIMEZONE
TZID:America/New_York
X-LIC-LOCATION:America/New_York
BEGIN:STANDARD
DTSTART:20251102T010000
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
END:STANDARD
END:VTIMEZONE
END:VCALENDAR