Towards Redundancy-Free Sub-networks in Continual Learning

Published: 2026, Last Modified: 05 Nov 2025Pattern Recognit. 2026EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•Our method introduces an information bottleneck into the continual learning process.•Our IBM ensures the construction of redundancy-free sub-networks.•IBM promotes effective knowledge transfer for learning new tasks.•Experiments across multiple benchmarks reaches a new SOTA in continual learning.
Loading