Incremental learning systems must keep acquiring new knowledge without forgetting what they already know. That tension, known as the plasticity-stability dilemma, is a central obstacle for models that learn continuously. A new arXiv preprint introduces Co2In, a framework that addresses this problem through a bi-level memory architecture.
Co2In structures learning around three processes: knowledge encoding, consolidation, and integration. By separating these stages, the framework can manage new information while preserving previously learned patterns. The authors position this as a way to balance plasticity and stability more explicitly than typical incremental learning methods.
The paper's abstract does not provide experimental details or comparisons, so the claims are limited to the proposed design. Still, the framework offers a conceptual contribution to continual learning, where maintaining both adaptability and durability remains an open challenge.