Jointly held with Statistics and Actuarial Science
M3 3127
Speaker
Venu Veeravalli (University of Illinois at Urbana-Champaign)
Title
Out-of-Distribution Detection via Multiple Testing
Abstract
Out-of-Distribution (OOD) detection in machine learning refers to the problem of detecting whether the machine learning model's output can be trusted at inference time. This problem has been described qualitatively in the literature, and a number of ad hoc tests for OOD detection have been proposed. In this talk we outline a principled approach to the OOD detection problem, by first defining the problem through a hypothesis test that includes both the input distribution and the learning algorithm. Our definition provides insights for the construction of good tests for OOD detection. We then propose a multiple testing inspired procedure to systematically combine any number of different OOD test statistics using conformal p-values. Our approach allows us to provide strong guarantees on the probability of incorrectly classifying an in-distribution sample as OOD. In our experiments, we find that the tests proposed in prior work perform well in specific settings, but not uniformly well across different types of OOD instances. In contrast, our proposed method that combines multiple test statistics performs uniformly well across different datasets, neural networks and OOD instances. We will end the talk with a discussion of the application of the multiple testing approach to the problem of hallucination detection in LLMs.