Release the Potential of Memory Buffer in Continual Learning: A Dynamic System Perspective.

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Continual learning (CL) aims to learn from non-stationary data without knowledge forgetting.
• Memory-replay methods in CL often suffer from memory overfitting due to limited buffer diversity and hardness.
• Existing solutions for memory overfitting lack data diversity/hardness or are complex to train.

Purpose of the Study:

• To address memory overfitting in continual learning by proposing a novel memory transformation method.
• To enhance memory buffer diversity and hardness using a dynamic system perspective.
• To improve the efficiency and effectiveness of memory-replay approaches in CL.

Main Methods:

• Proposed a continuous and reversible memory transformation method viewed from a dynamic system perspective.
• Introduced an adversarial optimization objective to jointly train the CL model and memory transformer.
• Developed deterministic continuous memory transformer (DCMT) and stochastic continuous memory transformer (SCMT) for diverse memory generation.

Main Results:

• Significantly increased memory buffer diversity and hardness, reducing overfitting.
• Demonstrated memory efficiency without requiring data replication.
• Achieved substantial performance improvements compared to strong baselines in extensive experiments.

Conclusions:

• The proposed neural transformation approaches effectively enhance memory buffer quality for continual learning.
• Continuous and reversible memory transformation offers a promising direction for mitigating memory overfitting.
• The method is memory-efficient and provides significant performance gains over existing techniques.