Class-Incremental Learning via Knowledge Amalgamation

Catastrophic forgetting has been a significant problem hindering the deployment of deep learning algorithms in continual setting. Numerous methods have proposed to address catastrophic where an agent loses its generalization power old tasks while new tasks. We put forward alternative strategy handle with knowledge amalgamation (CFA), which learns student network from multiple heterogeneous teacher models specializing previous and can be applied current offline methods. The process is carried out single-head manner only selected number memorized samples no annotations. teachers students do not need share same structure, allowing adapted compact or sparse data representation. compare our method competitive baselines different strategies, demonstrating approach’s advantages. Source-code: github.com/Ivsucram/CFA .