时 间:2026年12月14日(星期一)16:00-17:00
主持人:复旦大学管理学院统计与数据科学系 蒋斐宇 副教授
地 点:史带楼302室
报告人:张琼 博士
穆罕默德·本·扎耶德人工智能大学(MBZUAI)
题 目:Tabular Foundation Models: A New Paradigm for Statistical Learning
摘 要:
Most statistical learning methods are fitted from scratch for each new dataset. Tabular foundation models suggest a fundamentally different workflow: train a model once across a distribution of data-generating mechanisms, then use the trained model to perform in-context inference on new datasets without task-specific optimization. By amortizing inference across many problems, these models offer the prospect of fast, reusable, and broadly applicable statistical procedures.
In this talk, I introduce tabular foundation models through a statistical and approximate Bayesian perspective, using my recent work to examine three questions. First, what enables these models to adapt to new datasets? I show that they can respond to both parametric and nonparametric structure, and that their performance depends critically on alignment between the pretraining prior and model architecture. Second, when should we trust them? Evaluations beyond predictive accuracy reveal important differences in uncertainty quantification and subgroup calibration, while repeated measurements can paradoxically harm predictions when the deployment protocol conflicts with assumptions learned during pretraining. Third, how far can the paradigm extend? I demonstrate that prior-data fitted networks can move beyond supervised prediction to problems including clustering with unknown cluster cardinality and sequential decision-making through contextual bandits.
Together, these results position tabular foundation models not as universal black-box predictors, but as a new way to construct statistical procedures: encode assumptions through pretraining distributions, amortize computation across datasets, and perform inference through a single forward pass. They also point toward a research agenda centered on understanding, evaluating, and designing foundation models for reliable statistical inference.
个人简介:
张琼博士现任穆罕默德·本·扎耶德人工智能大学(MBZUAI)助理教授。在加入MBZUAI之前,她曾任中国人民大学统计与大数据研究院助理教授。她于2015年获得中国科学技术大学少年班学院学士学位,2022年获加拿大英属哥伦比亚大学统计学博士学位。其主要研究方向包括表格基础模型,联邦学习等机器学习领域,相关成果已发表在JASA、JRSSB、JMLR、IEEE TIT等顶级期刊,以及ICLR、NeurIPS, ICCV等国际权威会议上。
复旦大学管理学院统计与数据科学系