Abstract
A reliable uncertainty estimation method is the foundation of many modernout-of-distribution (OOD) detectors, which are critical for safe deployments ofdeep learning models in the open world. In this work, we propose TULiP, atheoretically-driven post-hoc uncertainty estimator for OOD detection. Ourapproach considers a hypothetical perturbation applied to the network beforeconvergence. Based on linearized training dynamics, we bound the effect of suchperturbation, resulting in an uncertainty score computable by perturbing modelparameters. Ultimately, our approach computes uncertainty from a set of sampledpredictions. We visualize our bound on synthetic regression and classificationdatasets. Furthermore, we demonstrate the effectiveness of TULiP usinglarge-scale OOD detection benchmarks for image classification. Our methodexhibits state-of-the-art performance, particularly for near-distributionsamples.