演讲人： 金滢 [宾夕法尼亚大学]
时间： 10:30-12:00, Jul 24, 2025 (Thu)
地点：RM 1-222, FIT Building
内容：
Conformal prediction is a distribution-free uncertainty quantification framework for assessing the reliability of black-box AI models. Standard conformal prediction provides on-average (marginal) guarantees which, despite being useful, can be insufficient in decision-making processes that usually come with a selective nature.
In this talk, I introduce a conformal selection framework that offers selective inference capabilities for conformal prediction. We focus on applications where predictions from black-box models are used to shortlist unlabeled test samples whose unobserved outcomes satisfy a desired property. Starting from a pool of unlabeled test instances and a labeled calibration set, we compute p-values for testing each candidate’s outcome. Applying the Benjamini–Hochberg procedure yields a data-driven cutoff that selects only those candidates whose conformal p-values fall below the threshold. We show that this procedure provides distribution-free control of the false discovery rate (FDR) in finite-sample.
I will demonstrate how this approach can be used to (1) accelerate hit discovery in virtual screening for drug candidates, and (2) establish trustworthy filtering in large-language-model alignment. Finally, I will survey extensions to other selective-inference tasks, highlighting how selective conformal prediction can offer opportunities for trustworthy AI-driven decisions.
个人简介:
Ying Jin is currently an Assistant Professor in Statistics and Data Science at the Wharton School, University of Pennsylvania. Prior to that, she was a Wojcicki-Troper Postdoctoral Fellow at Harvard Data Science Initiative from 2024 to 2025, working with Professors José Zubizarreta and Marinka Zitnik at Harvard Medical School. She obtained her PhD in Statistics from Stanford University in 2024, advised by Professors Emmanuel Candès and Dominik Rothenhäusler. Her research centers around statistical uncertainty quantification for black-box AI models, generalizability, distributional robustness, causal inference, and their applications in biomedical discovery and human decisions.
