Machine Learning with Statistical Guarantees
Developing methods that bring formal statistical guarantees to machine learning models used in health, policy, and other high-stakes decisions.
Lead: Daniel Csillag, Tegoni Goedert
Machine learning models are increasingly used to inform decisions that matter — who gets flagged for follow-up care, which patients are at risk, what a policy change is likely to do. But a model's predictions are only as trustworthy as the uncertainty attached to them, and most ML methods offer no formal guarantee about how often, or how badly, they can be wrong.
This project develops statistical methods that close that gap: conformal prediction techniques that produce valid uncertainty estimates regardless of the underlying model, generalization bounds for causal regression that make explicit what a model's estimated effect actually promises, and prediction-powered inference methods that let researchers combine cheap, noisy model outputs with small amounts of high-quality labeled data without losing statistical validity. The goal throughout is the same: keep the predictive power of modern machine learning while making its guarantees explicit and testable, rather than assumed.
Much of this work is developed by Daniel Csillag, whose research under this project has been published at leading machine learning venues including ICML, NeurIPS, and AISTATS. The methods developed here feed directly into LABORA's applied health projects, wherever a model's output needs to be trusted enough to act on.