The article by Anish Lakkapragada, Amol Khanna, Edward Raff, and Nathan Inkawhich discusses the importance of detecting when data is outside the expected input distribution in machine learning. It presents a new method, DICE-COL, which modifies the popular Directed Sparsification (DICE) method, resolving a flaw in its design. The authors also explore dimensionality reduction for representation-based methods in OOD detection. Their methods demonstrated significant performance improvements on the OpenOODv1.5 benchmark framework.
Publication date: 18 Jan 2024
Project Page: https://arxiv.org/abs/2401.10176v1
Paper: https://arxiv.org/pdf/2401.10176
