Class-incremental learning with causal relational replay

Toan Nguyen, Duc Kieu, Bao Duong, Tung Kieu, Kien Do, Thin Nguyen, Bac Le

Abstract: Highlights•We tackle the problem of catastrophic forgetting in Class-IL.•Our method leverages object relations, inspired by human memory, to prevent forgetting.•We discover spurious relations worsen forgetting, best to focus only on intrinsic ones.•We add a causal learning mechanism to analyze relations and combat spurious factors.•Our algorithm excels in Class-IL, consistently surpassing top rehearsal techniques.