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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Class-incremental learning (CIL) aims to learn new tasks without forgetting previous ones.
• Dynamic task sequences present challenges like unclear category boundaries and context bias in CIL.
• Existing CIL methods often struggle with effective knowledge retention and adaptation.

Purpose of the Study:

• To propose a novel class-incremental learning approach that supports causal reasoning and mitigates forgetting.
• To introduce the Causal Inference Framework for CIL (CAFF-CIL) to address limitations in current CIL methods.
• To enhance model performance in dynamic learning environments by improving category discriminative boundaries and reducing context bias.

Main Methods:

• The proposed CAFF-CIL framework integrates three key components: Task-Adaptive Causal Feature Selection (TAFS), Causal Dual-Path Modulation (CDPM), and Task-Adaptive Hyperparameter Tuning (TAHT).
• TAFS quantifies feature causality and creates a forgetting channel for redundant features.
• CDPM stabilizes base class representations while adapting novel class features, supported by TAHT's stage-aware optimization strategy.

Main Results:

• CAFF-CIL demonstrated a reduction in the forgetting rate by 7.91%.
• The framework achieved an accuracy improvement of 4.8% on the CIFAR-100 dataset.
• Experiments across eight benchmark datasets validated the effectiveness of the proposed approach.

Conclusions:

• The CAFF-CIL framework effectively addresses forgetting and context bias in class-incremental learning through causal reasoning.
• The integration of TAFS, CDPM, and TAHT components leads to significant performance gains in dynamic learning scenarios.
• This causal inference approach offers a promising direction for advancing robust and efficient class-incremental learning systems.