[3Xin4] Poster session 1

[3Xin4-23] Performance Evaluation and Analysis of Out-of-Distribution Detection in Continual Learning

〇Tomoya Kaichi¹, Tatsuya Konishi¹, Wataru Kumagai^1,2, Mori Kurokawa¹ (1.KDDI Research, Inc., 2.The University of Tokyo)

Keywords:OOD detection, Continual learning

Deep learning systems typically suffer from catastrophic forgetting of past knowledge when learning new tasks incrementally. Continual learning (CL) is a paradigm that aims to overcome the forgetting. When a model trained in CL setting is used in the real-world environment, the ability to identify out-of-distribution (OOD) samples ensures that the model is safe and reliable when making decisions. Furthermore, OOD detection in CL scenario would progress to open-world learning, where an intelligent system identifies novel tasks and learns them incrementally. In this paper, we investigate the relation with the forgetting and the accuracy of OOD detection in CL. Through the extensive experiments, we found that forgetting does not necessarily correlate with deterioration of the OOD detection performance, which implies that OOD detection in CL requires other efforts in addition to improving the accuracy of CL.