Locally adaptive compensation for analytic class-incremental learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Class-Incremental Learning (CIL) faces challenges like catastrophic forgetting and the stability-plasticity dilemma due to lack of past data.
• Analytic CIL (ACIL) methods use closed-form updates but can suffer from underfitting and limited representational flexibility.

Purpose of the Study:

• To introduce Locally Adaptive Compensation Learning (LACL), an analytic framework designed to overcome limitations in existing ACIL methods.
• To enhance representational robustness and manage the stability-plasticity trade-off in CIL.

Main Methods:

• LACL employs a neighborhood-aware compensation mechanism utilizing local representations.
• The framework is formulated using a closed-form recursive approach, ensuring interpretability and theoretical rigor.
• This method is an exemplar-free approach, meaning it does not require storing past data samples.

Main Results:

• LACL demonstrates state-of-the-art performance among exemplar-free CIL methods.
• The proposed method shows increasing advantages over prior approaches with a growing number of incremental phases.
• Experiments conducted on CIFAR-100, ImageNet-100, and ImageNet-Full validate the effectiveness of LACL.

Conclusions:

• LACL effectively reduces underfitting and improves the stability-plasticity balance in CIL.
• The neighborhood-aware compensation mechanism strengthens representational robustness.
• LACL offers a promising, interpretable, and theoretically sound solution for challenging CIL scenarios.