讲座标题(中文):f-散度在现代学习理论中的统一视角:从领域自适应到弱到强泛化
讲座标题(英文): A Tour of f-Divergence in Modern Learning Theory: From Domain Adaptation to Weak-to-Strong Generalization
讲座摘要:The family of f-divergences provides a versatile language for measuring distribution shift and statistical dependence, with applications spanning domain adaptation, generalization theory, and modern large language model training. In this talk, I will introduce the basic variational perspective of f-divergence and show how it serves as a common tool across three seemingly different problems. First, I will discuss an improved f-divergence framework for unsupervised domain adaptation, where refined discrepancy measures lead to sharper target-risk and generalization guarantees. Second, I will present conditional f-information bounds, which extend conditional mutual information methods and obtain new information-theoretic generalization bounds in the supersample setting. Finally, I will turn to weak-to-strong generalization in LLMs, where f-divergence-based losses provide a principled way to analyze and improve learning from weak supervision, while also revealing fundamental limitations and equivalences among different losses.
讲者信息:汪子乔现为同济大学计算机科学与技术学院助理教授、博士生导师,深研究方向为机器学习基础理论、大模型学习理论与算法以及信息论。近几年主要成果发表在人工智能、机器学习及数据挖掘等相关领域国际顶级会议,涵盖NeurIPS、ICML、ICLR、UAI、AAAI、KDD、WWW等,博士论文被提名2025年加拿大人工智能协会最佳博士论文奖,以及提名2025年渥太华大学总督学术奖章和Pierre Laberge论文奖。目前担任ICLR, NeurIPS等人工智能会议的领域主席,曾担任2024年IEEE 北美信息论暑期学校(NASIT)联合程序主席。个人主页:https://ziqiaowanggeothe.github.io/