Hoi-To Wai @ CUHK

Associate Professor Department of Systems Engineering and Engineering Management The Chinese University of Hong Kong, Hong Kong Email: htwai at se dot cuhk dot edu dot hk Links: My CV, Department Profile, Google Scholar

About Me

I am an Associate Professor at the Department of Systems Engineering and Engineering Management in The Chinese University of Hong Kong (CUHK). I also received my B. Eng (with first class Honor) and M. Phil. from CUHK, working with Prof. Wing-kin (Ken) Ma in Spring 2010 and 2012, respectively. I received my PhD degree from Arizona State University in Fall 2017, working with Prof. Anna Scaglione. Currently, I am serving the editorial board of IEEE TSP and Elsevier's Signal Processing, as well as a member of the IEEE SPS's SPTM-TC and DSI. I am also an organizer for the DEGAS Webinar Series sponsored by the IEEE SPS.

Besides ASU and CUHK, I have held research positions at LIDS, MIT working with Prof. Asuman E. Ozdaglar, and at Télécom ParisTech, Ecole Polytechnique, working with Prof. Eric Moulines, and at UC Davis as a PhD student with Anna before transferring to ASU. I have received the 2017's Dean's Dissertation Award at ASU's Ira A. Fulton Schools of Engineering (Full Circle), and a Best Student Paper Award from IEEE ICASSP 2018 (paper), IEEE SAM Workshop 2024 (as co-author, paper), IEEE ICASSP 2025 (as co-author, paper).

My current research focus is on:

• Graph signal processing and graph learning problem for network science

• Stochastic algorithms for machine learning and distributed signal processing

Check out the Research page for more details.

I am looking for mathematically inclined and motivated students/Postdoc interested in optimization theory for machine learning, reinforcement learning or graph signal processing problems. Please contact me if you are interested. Also consider applying for the HK PhD Fellowship Scheme which offers great support for studying in CUHK.

Recent News

• Nov, 2025: Two papers accepted by IEEE Transactions on Signal Processing:

• • Decentralized Stochastic Optimization over Unreliable Networks via Two-timescales Updates (with Haoming & Oscar)  —  we proposed a systematic way to design compression-enabled decentralized optimization algorithms for unreliable networks. Also see the slides for a recent talk I have given at NUS.

• • Multilinear Tensor Low-Rank Approximation for Policy-Gradient Methods in Reinforcement Learning (with Sergio & Antonio)  —  We showed how to efficiently learn a tensor-factored policy model.

• Nov, 2025: I have received a Dean Exemplary Teaching Award 2024 from the Faculty of Engineering, CUHK.

• Oct., 2025: I have visited UC Santa Barbara and gave a talk on Network Games Induced Prior for Graph Topology Learning. Thanks Mahnoosh for hosting me.

• August, 2025: Three of my PhD students have defeneded their thesis: Qiang is going to Huawei, Chenyue will be a postdoc at Cornell Tech, and Oscar will be a postdoc at U of Minnesota. Congratulations!

• July, 2025: I have visited UC Irvine and gave a talk on Network Games Induced Prior for Graph Topology Learning. Thanks Yanning for hosting me.

• July, 2025: 2 papers accepted at IEEE CDC 2025:

• • Asynchronous and Stochastic Distributed Resource Allocation (with Qiang and Michal)  —  We showed that primal-dual-based distributed optimization can be accelerated using asynchronous & stochastic updates.

• • Stochastic Gradient Descent with Strategic Querying (with Nanfei and Mahnoosh)  —  We showed that the transient performance of SGD can be accelerated by querying the component function strategically. Most notably, the performance gain can be analyzed and predicted.

• May, 2025: 2 papers on performative prediction and Quantized LLM Training accepted at ICML 2025:

• • Clipped SGD Algorithms for Performative Prediction: Tight Bounds for Stochastic Bias and Remedies (with Qiang & Michal)  —  We showed clipping SGD displays an intriguing bias amplification behavior under the performative prediction setting. Particularly, the more sensitive the distribution shift is, the more bias will be found. We also suggest remedies through step size tuning or deploying a error-feedback inspired algorithm.

• • RoSTE: An Efficient Quantization-Aware Supervised Fine-Tuning Approach for Large Language Models (with Wei, Oscar, and others)  —  We show that “rotation + STE” enables one to achieve SOTA performance in LLM fine tuning.