[108th TrustML Young Scientist Seminar] Talk by Yu-Jie Zhang (University of Washington) "The Role of Curvature in Non-stationary Online Learning"
RIKEN AIP seminar on how loss curvature speeds up non-stationary online learning, by Yu-Jie Zhang (University of Washington).
- When
- Where
- Meeting Room C, RIKEN Nihonbashi Office, Tokyo·Hybrid
- Region
- Kanto (Tokyo)
- Organizer
- RIKEN Center for Advanced Intelligence Project
- Language
- EN
- Source
- Doorkeeper
![[108th TrustML Young Scientist Seminar] Talk by Yu-Jie Zhang (University of Washington) "The Role of Curvature in Non-stationary Online Learning"](https://doorkeeper.jp/rails/active_storage/representations/proxy/eyJfcmFpbHMiOnsibWVzc2FnZSI6IkJBaHBBMVYyQVE9PSIsImV4cCI6bnVsbCwicHVyIjoiYmxvYl9pZCJ9fQ==--a5c843cb8358ea510f8b99f14ae64db5d267f999/eyJfcmFpbHMiOnsibWVzc2FnZSI6IkJBaDdCem9MWm05eWJXRjBTU0lJYW5CbkJqb0dSVlE2RTNKbGMybDZaVjloYm1SZmNHRmtXd2RwQWNocEFjZz0iLCJleHAiOm51bGwsInB1ciI6InZhcmlhdGlvbiJ9fQ==--ed21893bc3f6dbb42a45b555f9878f5434977670/aip_logo_w200.jpg)
Summary
Yu-Jie Zhang of the University of Washington presents research on online learning, where decisions are made sequentially from incoming feedback. The talk asks when and how loss curvature can speed up learning if the best decision changes over time.
He shows that with bounded gradients, curvature accelerates learning on arbitrary bounded convex domains without the out-of-domain predictions earlier work required. With unbounded gradients, as in online portfolio selection, faster rates depend on the geometry of the changing comparators, particularly their position relative to the boundary. The results come from dynamic regret analysis with new mixability-based arguments.
The seminar is held at the RIKEN Nihonbashi Office and online. In-person attendance is limited to AIP researchers. It suits researchers interested in the theory of online optimization, bandits and reinforcement learning.
About the community
The TrustML Young Scientist Seminar is a series hosted by RIKEN AIP featuring early-career researchers presenting work on trustworthy machine learning and its mathematical foundations.
#machine-learning#online-learning#optimization#research-seminar#riken#theory