BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//IFDS - ECPv6.0.1.1//NONSGML v1.0//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:IFDS
X-ORIGINAL-URL:https://ifds.info
X-WR-CALDESC:Events for IFDS
REFRESH-INTERVAL;VALUE=DURATION:PT1H
X-Robots-Tag:noindex
X-PUBLISHED-TTL:PT1H
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:20210314T100000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:20211107T090000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTART;TZID=America/Los_Angeles:20210402T133000
DTEND;TZID=America/Los_Angeles:20210402T143000
DTSTAMP:20260516T085406
CREATED:20210115T201436Z
LAST-MODIFIED:20210331T134630Z
UID:880-1617370200-1617373800@ifds.info
SUMMARY:IFDS All-Hands: Chaobing Song & Xuezhou Zhang
DESCRIPTION:Speaker 1: Dr. Chaobing Song\, IFDS Postdoctoral Scholar at U Wisconsin (advised by Prof.s Jelena Diakonikolas and Steve Wright)\n\nTitle: Closing Convergence Gaps for Both Smooth and Nonsmooth Convex Finite-Sums\n \nAbstract: In the Foundation of Data Science\, optimization problems with a finite-sum structure widely exist\, such as the classical empirical risk minimization problem and deep neural network. In the past decade\, one main concern in the study of first-order methods is to study how the finite-sum structure influences optimization efficiency and scalability. In this talk\, based on my work (https://arxiv.org/abs/2006.10281\, https://arxiv.org/abs/2102.13643)\, I will talk about a consistent approach based on dual averaging and a particularly designed initialization strategy to close convergence gaps for smooth convex finite-sums and nonsmooth convex finite-sums. For the first time\, the proposed VRADA algorithm (https://arxiv.org/abs/2006.10281) matches the lower bounds of all the three regimes for smooth convex finite-sums; for the first time\, the proposed VRPDA^2 algorithm  (https://arxiv.org/abs/2102.13643) shows an O(n) improvement theoretically over existing deterministic methods and stochastic primal-dual coordinate methods\, where n is the number of data samples. Both algorithms also have good empirical performance.\n\n \nSpeaker 2: Xuezhou Zhang\, IFDS RA (advised by Prof. Jerry Zhu)\n \n\nTitle: Statistical Robustness in Reinforcement Learning\n \nAbstract: Traditional robust statistics studies the problem of statistical estimation under data contamination. In this talk\, we extend this investigation to decision-making problems. In the first half of the talk\, I will illustrate the concept and unique challenge of robust Decision making using a multi-armed bandit as an example. In the second half of the talk\, I will present some of the lower and upper-bound results we have already established\, and discuss several open directions. This talk is partially based on our recent paper [https://arxiv.org/pdf/2102.05800.pdf].\n\n \n\n—-\nSpeaker Bios:\n \nDr. Chaobing Song: I am a postdoc researcher at University of Wisconsin-Madison with Prof. Jelena Diakonikolas and Prof. Stephen J Wright. My research interests include optimization and machine learning. I obtained my Ph.D. degree at Tsinghua University in 2020 supervised by Prof. Yi Ma from University of California\, Berkeley (who is an affiliate professor at Tsinghua-Berkeley Shenzhen Institute\, Tsinghua University). I spent two wonderful years at Berkeley and finished my Ph.D. thesis there. The main focus of my current research is to apply a general\, powerful and concise framework [arXiv] to optimization problems in machine learning. The development based on this framework substantially improves complexity results in many well-known settings\, such as the setting of nonsmooth convex finite-sums in this talk.  \n \nXuezhou Zhang: Xuezhou is a Ph.D. candidate in the computer sciences department at the University of Wisconsin-Madison\, advised by Professor Jerry Zhu. Before coming to Madison\, he obtained a Bachelor of Applied Mathematics degree from UCLA. His research interests lie in machine learning and reinforcement learning\, and his recent research focuses on designing reinforcement learning algorithms that learn more adaptively and robustly in non-stochastic environments.\n\nThese talks are remote via zoom. Please contact the organizer if you need the link. \nAll-Hands titles and abstracts are tentative\, as of the posting date.
URL:https://ifds.info/event/ifds-all-hands-040221/
LOCATION:WI
CATEGORIES:Monthly All-Hands
END:VEVENT
END:VCALENDAR