Huaiguang Cai 蔡怀广
 |
Independent Researcher
|
Research Interests
My research aims to develop reliable, explainable, and practical decision-making algorithms. I am also interested in methods with elegant mathematical modeling. Currently, my focus is on enhancing the reasoning capabilities of large language models.
Algorithms (Algorithmic Game Theory, Online Learning Theory, Ad Auction)
Machine Learning (Explainable AI, Edge computing, Recommendation system)
Large Language Models (Long Context, Enhanced Reasoning, Efficient Training)
Publications
Online Resource Allocation for Edge Intelligence with Colocated Model Retraining and Inference
[Paper]
[Slides]
[Code]
Huaiguang Cai, Zhi Zhou (Sun Yat-sen University), Qianyi Huang (Sun Yat-sen University)
In IEEE International Conference on Computer Communications (INFOCOM, CCF-A), Vancouver, Canada, 2024
CHG Shapley: Efficient Data Valuation and Selection towards Trustworthy Machine Learning
[OpenReview]
[Code1]
[Code2]
Huaiguang Cai
Arxiv
CAMs as Shapley Value-based Explainers
[Paper]
[Code]
[Merged into the official GradCAM repository (11k stars)]
Huaiguang Cai
The Visual Computer (SCI)
Project
Learning in games, last-iterate, Blackwell approachability theorem
[Results]
Experimentally observed that predictive CFR algorithms converge much faster in last-iterate than in average-iterate in non-HUNL two-player zero-sum games.
I am now working on connecting the Proportional-Integral-Derivative (PID) algorithm in control theory with the Counterfactual Regret Minimization (CFR) algorithm in game theory by developing a last-iterate version of the Blackwell approachability theorem.
Education Background
I am a master's student majoring in Pattern Recognition and Intelligent Systems at Institute of Automation, Chinese Academy of Sciences (CASIA), University of Chinese Academy of Sciences (UCAS), advised by Prof. Wensheng Zhang.
I received my bachelor's degree in Computer Science from School of Computer Science and Engineering, Sun Yat-Sen (Zhongshan) University, advised fortunately by Prof. Zhi Zhou and Prof. Chang-Dong Wang.
