蔣兆凱 / Chao-Kai Chiang
I develop principled machine learning methods for imperfect supervision, noisy labels, online learning, and bandit feedback.
My work connects theory, algorithms, and practical systems for learning when labels are weak or unreliable, distributions shift, or feedback is partial and sequential.
Recruiting
I welcome motivated students interested in principled machine learning, especially weak supervision, noisy labels, distribution shift, online learning, bandit algorithms, and adaptive learning systems. Prospective students are encouraged to email me with a CV and a brief description of their research interests.
Research Vision
Modern machine learning is increasingly deployed in settings where clean labels, stable distributions, and complete feedback are unavailable. My research aims to build a principled foundation for learning under such realistic conditions: weak or noisy supervision, distribution shift, sequential feedback, and adaptive decision making. The long-term goal is to understand the weakest yet learnable forms of supervision and to design algorithms that remain reliable beyond idealized data assumptions.
Research
My current research develops theory and algorithms for machine learning systems that must learn from imperfect supervision, changing environments, and partial feedback. I focus on problems where statistical reliability, optimization behavior, and practical applicability must be studied together.
I study weakly supervised learning and learning with label noise, including unified risk analysis, reliable validation under noisy labels, robust imitation learning from vague feedback, and learning under complex supervision transitions.
I develop algorithms and regret analyses for online convex optimization and multi-armed bandits, with interests in best-of-both-worlds guarantees, Thompson sampling, Pareto front identification, and contextual or dueling feedback.
I connect theoretical insights with practical systems, including federated multi-task learning, budgeted hyper-parameter tuning, LLM routing, and adaptive decision making with preference or bandit feedback.
Selected Publications
Selected first by recent top-conference publications, then by citation impact within the main research areas.
A unified contamination-decontamination view of weak supervision and risk rewriting across WSL/LLN settings.
Provides an optimal-transport view of domain adaptation and introduces entanglement as a term explaining when alignment fails.
Studies how prior choices affect Thompson sampling in multiparameter bandit models.
Applies weakly supervised learning ideas to recover useful imitation signals from vague feedback.
Analyzes when Thompson sampling is optimal for heavy-tailed Pareto bandits and how truncation restores optimality.
A systems-aware federated multi-task learning framework for heterogeneous distributed data.
Introduced variation-aware regret analysis for online optimization in gradually changing environments.
A best-of-both-worlds bandit algorithm with nearly optimal pseudo-regret across stochastic and adversarial settings.
Publications
The list below is compiled from the uploaded publication list and CV. Use the filters to browse by category.
Transactions on Machine Learning Research, 2025.
Journal of Information Hiding and Multimedia Signal Processing 8(4): 869-879.
International Conference on Artificial Intelligence and Statistics, pp. 3034-3042.
AAAI Conference on Artificial Intelligence, pp. 13383-13390.
Advances in Neural Information Processing Systems, pp. 48275-48292.
International Conference on Machine Learning, pp. 18810-18851.
International Joint Conference on Artificial Intelligence, pp. 5744-5750.
Advances in Neural Information Processing Systems, pp. 4424-4434.
Annual Conference on Learning Theory, pp. 116-120. Alphabetical order; full version available as preprint.
International Conference on Artificial Intelligence and Statistics, pp. 939-947.
International Conference on Robot, Vision and Signal Processing, pp. 191-194. Short conference paper.
Asian Conference on Machine Learning, pp. 344-359.
Annual Conference on Learning Theory, pp. 210-227. Alphabetical order.
Annual Conference on Learning Theory, pp. 6.1-6.20. Mark Fulk Best Student Paper Award.
Annual ACM-SIAM Symposium on Discrete Algorithms, pp. 616-629.
arXiv:2510.00841. Under double-blind review in uploaded publication list.
arXiv:2302.14407v1. Preliminary version of the AAAI 2024 paper.
arXiv:1902.00532. Preliminary version of the IJCAI 2019 paper.
arXiv:1605.08722. Full version of the COLT 2016 paper.
Contact
Department of Computer Science
National Yang Ming Chiao Tung University