CAFF-CIL: Causality-Aware Freedom Forgetting Approach for Class-Incremental Learning.

Abstract

Class-incremental learning (CIL) requires free learning of all previously learned tasks. This is desirable to achieve without task indexing when facing dynamic task sequences. However, problems such as unclear category discriminative boundaries and context bias are often faced. To address these issues, we propose a freedom forgetting CIL approach that supports causal reasoning. The causal inference framework for CIL (CAFF-CIL) framework consists of three key components: 1) the task-adaptive causal feature selection (TAFS) component, which quantifies the causal contribution of features to classification decisions and constructs a freedom forgetting channel for redundant features; 2) the causal dual-path modulation (CDPM) component, which maintains the stable causal representation of base classes while achieving causal adaptation of novel class features; and 3) the task-adaptive hyperparameter tuning (TAHT) component, which leverages a stage-aware scheduling strategy to adaptively adjust optimization priorities across different incremental stages. Extensive experiments on eight benchmark datasets verify that CAFF-CIL reduces the forgetting rate by 7.91% and improves the accuracy by 4.8% on CIFAR-100.