Advertisement
🧠 New Seminar 學者開講 | HKU IDS Scholar Seminar Series #29Large-scale nonconvex optimization becomes especially challenging under stochasticity, where conventional stochastic first-order methods can suffer from sample complexity that grows linearly with problem dimension.
In this seminar, Dr Yue Xie will present recent approaches to tackling these challenges through non-Euclidean optimization and parameter-free stochastic methods.
The first part of the talk explores the use of non-smooth proximal terms in stochastic gradient steps under non-Euclidean settings. This can lead to stronger convergence metrics, limited additional computational cost, and potentially dimension-insensitive sample complexity. Dr Xie will also discuss variance-reduction techniques that achieve near-optimal sample complexity, including what is believed to be the first such result in the ℓ₁/ℓ∞ setting.
The second part introduces an Armijo-enabled stochastic line-search framework that requires no prior knowledge of the Lipschitz smoothness constant or other problem parameters. The framework also accommodates a simple nonsmooth convex component through proximal-gradient updates, with corresponding guarantees under Polyak–Łojasiewicz and convex settings.
🔗 Details and registration: https://datascience.hku.hk/s261014f
🗣️ Dr Yue Xie is a Member of HKU IDS and a Research Assistant Professor at HKU Math. His research focuses on the design and analysis of algorithms for nonconvex and stochastic optimisation, with applications in machine learning and data science. He previously worked as a postdoctoral researcher at the University of Wisconsin–Madison in the nonconvex optimisation group led by Prof Stephen J. Wright, and received his PhD from Pennsylvania State University and bachelor’s degree from Tsinghua University.
‼️ Seats for on-site participation are limited. Register now and join us in person or on Zoom.
今期 HKU IDS 學者講座，謝越博士將介紹兩類處理非凸隨機優化問題的方法，包括非歐幾里得優化及無需預先設定問題參數的隨機優化方法。
講座將探討如何透過非光滑近端項及變異數縮減方法，降低高維問題中的樣本複雜度，並介紹一個結合 Armijo 準則的隨機線搜索框架，在毋須預先知道 Lipschitz 光滑常數或其他問題參數的情況下，仍能提供相應的收斂保證。
🔗 詳情及報名：https://datascience.hku.hk/s261014f
#HKU #HKUIDS #DataScience #Optimization
Advertisement
Event venue
The University of Hong Kong, Stephen Hui Geological Museum, Sai Wan, Hong Kong, Hong Kong SAR
Tickets
Concerts, fests, parties, meetups — all the happenings, one place.
Discover more events by tag
Conferences in Hong KongEducation in Hong KongTechnology in Hong KongArtificial-intelligence in Hong KongData-science in Hong KongMachine-learning in Hong KongWorkshops in Hong KongScience-fair in Hong Kong