Revisiting Flatness-Aware Optimization in Continual Learning With Orthogonal Gradient Projection.

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Continual learning (CL) aims to learn new tasks without forgetting previous ones.
• Orthogonal gradient projection (OGP) methods prevent forgetting but limit new task performance.
• Existing OGP methods face a trade-off between new task learning and catastrophic forgetting.

Purpose of the Study:

• To address the limitations of OGP-based CL methods.
• To propose a novel framework that enhances performance on new tasks while mitigating forgetting.
• To investigate the relationship between loss landscape flatness and catastrophic forgetting.

Main Methods:

• Established a unified framework for OGP-based CL methods.
• Analyzed OGP methods through the lens of loss landscape flatness.
• Proposed a Dual Flatness-aware OGD framework optimizing loss landscape flatness at data and weight levels.
• Implemented data and weight perturbation, flatness-aware optimization, and gradient projection modules.

Main Results:

• The proposed framework improves loss landscape flatness.
• Enhanced performance on new tasks is achieved.
• State-of-the-art (SOTA) average accuracy across all tasks was obtained.
• Demonstrated mitigation of catastrophic forgetting in continual learning.

Conclusions:

• Optimizing loss landscape flatness is crucial for effective continual learning.
• The Dual Flatness-aware OGD framework offers a promising solution to the catastrophic forgetting problem.
• This approach balances learning new information with retaining old knowledge effectively.