Using samples with label noise for robust continual learning

Hongyi Nie, Shiqi Fan, Yang Liu, Quanming Yao, Zhen Wang

Abstract: Highlights•We find that the label shift problem in continual machine learning breaks the label space consistency premise of the existing noise-using method, and it causes these methods not to provide reliable label estimations or loss correction for samples with label noise.•We propose SANU, a shift-adaptive noise-using method, which applies attention-based meta-knowledge techniques to help detect the label spaces of samples with label noise and pseudo-learning strategies to use samples with label noise from different sources effectively.