报告题目：Statistical Learning Theory for Contrastive Learning and LLM Alignment

报告人：Professor Yiming Ying，University of Sydney,Australia

报告时间：2026年7月17日（周五）9:30-12:30

报告地点：20-200

报告摘要：Algorithms in machine learning and AI do critically depend on at least three key components: (i) the risk function, which is the expectation of the loss function, (ii) the function space, which is often called the hypothesis space, and (iii) the set of probability measures, which are allowed for the specified algorithm. This talk gives a survey of a certain class of loss functions, which we call ratio-based. In supervised learning, margin-based loss functions for classification problems depending on the product of the output values and the predictions as well as distance-based loss functions depending on the difference of the output Modern AI changes the learning objects, but many of the underlying statistical questions remain classical. This talk revisits the framework of statistical learning theory—population targets, surrogate calibration, generalization, and approximation—through two examples: contrastive representation learning and large language model alignment. For contrastive learning, we identify the Bayes retrieval score, establish Fisher consistency and calibration of the contrastive logistic loss, and examine how negative sampling and representation approximation affect performance. For Direct Preference Optimization, we introduce contextual preference accuracy as a population alignment target and establish calibration of the DPO logistic objective. We then show that the score-difference parameterization imposes the Bradley–Terry structure: exact representation is equivalent to vanishing triangle curl, while nonzero curl captures an irreducible cyclic component of human preferences. These results illustrate how classical learning theory can clarify both the success and limitations of modern AI objectives.

报告人简介：Dr. Ying is a professor in the School of Mathematics and Statistics at the University of Sydney. Previously, he was a tenured professor in the Department of Mathematics and Statistics at SUNY Albany (USA), where he was also affiliated with Computer Science and founded the UAlbany Machine Learning Group. Dr. Ying received the University at Albany’s Presidential Award for Excellence in Research and Creative Activities (2022) and the SUNY Chancellor’s Award for Excellence in Scholarship and Creative Activities (2023). He regularly serves as aSenior Area Chair for major machine learning conferences such as NeurIPS, ICML, AAAI, and AISTATS. His research focuses on the theory and algorithms of machine learning and deep learning, with active collaboration with IBM on trustworthy AI.

邀请人：向道红