Prototypes as Anchors: Tackling Unseen Noise for online continual learning

Open Webpage

Shaoyuan Li, Yuxiang Zheng, Sheng-Jun Huang, Songcan Chen, Kangkan Wang

Published: 01 Jan 2025, Last Modified: 31 Jul 2025Neural Networks 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•We define closed-set and open-set noise in online CIL, addressing unseen classes overlooked by prior work.•We introduce PAA, a replay method that learns prototypes and uses similarity denoising for unseen classes.•Our experiments show PAA outperforms SOTA methods on synthetic and real noise benchmarks.