PASS++: A Dual Bias Reduction Framework for Non-Exemplar Class-Incremental Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Class-incremental learning (CIL) enables models to learn new classes sequentially.
• Existing CIL methods often rely on storing old data (exemplars), leading to catastrophic forgetting when data is unavailable.
• Catastrophic forgetting stems from representation and classifier biases in CIL models.

Purpose of the Study:

• To propose a novel non-exemplar class-incremental learning framework.
• To address representation and classifier biases inherent in CIL.
• To develop a method that mitigates forgetting without storing old data.

Main Methods:

• A dual-bias reduction framework combining self-supervised transformation (SST) and prototype augmentation (protoAug).
• SST enhances representation generalization by learning diverse, task-agnostic features.
• protoAug strengthens old class prototypes in feature space to preserve decision boundaries.

Main Results:

• The proposed framework significantly reduces forgetting in CIL.
• Performance is comparable to state-of-the-art exemplar-based methods, despite not storing old data.
• The method integrates seamlessly with pre-trained models.

Conclusions:

• Rethinking the necessity of storing old samples is crucial for advancing non-exemplar CIL.
• The dual-bias reduction framework offers an effective solution for data-efficient CIL.
• This work encourages further research into exemplar-free CIL methods.