Localized Conformal Inference for Operator Models

Presented by: Trevor Harris, Assistant Professor, Department of Statistics, University of Connecticut

Date: Wednesday, April 30, 2025, 4:00 PM, AUST 202

Meeting Link: Link

Coffee will be served at 3:30 in the Noether Lounge (AUST 326)

Abstract:

Operator models are regression algorithms between spaces of functions and have become key tools for learning large-scale dynamical systems. Recent advances in deep neural operators have significantly improved the accuracy and scalability of operator modeling. However, these models lack a built-in mechanism for predictive uncertainty quantification (UQ). In this work, we propose a general framework for UQ in operator models based on randomized local conformal inference and infimum depths. We show that our method achieves exact marginal coverage across a broad range of localizers, depth notions, and projection classes. Additionally, we introduce a sampling scheme to approximate prediction sets in real space and show that these sets are highly adaptive and effectively control coverage risk in both synthetic and real-world scenarios. Compared to existing baseline UQ methods for operator models, our approach yields tighter coverage and narrower, more adaptive prediction sets that better respond to variations in the residual distribution.

Bio: Trevor Harris is an assistant professor at the University of Connecticut in the Department of Statistics. He was previously an assistant professor at Texas A&M University before joining UConn in 2024. His research is in scientific machine learning, black box uncertainty quantification, operator modeling and functional data analysis, and causal impact analysis. Current projects include non-exchangeable conformal inference, climate model emulation and integration with observational data, large scale granger causal testing, and forecasting West Nile virus. Past research includes work in functional anomaly detection, change point detection, and robust nonparametric hypothesis testing.