Distributionally Robust Memory Evolution With Generalized Divergence for Continual Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Continual learning (CL) models struggle with non-stationary data distributions and knowledge forgetting.
• Existing memory replay methods in CL can overfit stored data, limiting generalization.
• Current CL approaches often ignore the uncertainty within memory data distributions.

Purpose of the Study:

• To propose a principled memory evolution framework for continual learning.
• To dynamically adjust memory data distributions to improve model performance.
• To enhance robustness against adversarial examples in CL settings.

Main Methods:

• Employed distributionally robust optimization (DRO) to dynamically evolve memory buffer distributions.
• Utilized f-divergence and Wasserstein ball constraints within DRO.
• Derived Wasserstein gradient flow (WGF) methods for f-divergence constraints.
• Developed direct solutions for Wasserstein ball constraints in Euclidean space.

Main Results:

• Demonstrated significant alleviation of forgetting in extensive benchmark experiments.
• Showcased improved generalization capabilities compared to existing CL methods.
• Observed enhanced robustness to adversarial examples as a byproduct of the framework.

Conclusions:

• The proposed memory evolution framework effectively addresses forgetting in continual learning.
• Distributionally robust optimization offers a principled approach to managing memory data.
• The framework provides a dual benefit of improved learning and adversarial robustness.