PVBF: A framework for mitigating parameter variation imbalance in online continual learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Online continual learning (OCL) enables AI to adapt to changing data streams.
• Experience replay (ER) methods store past data but can cause prediction bias due to parameter update deviations.
• Parameter variation imbalance is identified as a key cause of prediction bias in ER-based OCL.

Purpose of the Study:

• To identify and address parameter variation imbalance in ER-based OCL.
• To propose a novel framework, Parameter Variation Balancing Framework (PVBF), to mitigate prediction bias.
• To enhance the adaptive learning capabilities of AI systems in non-stationary environments.

Main Methods:

• Developed a method to evaluate parameter variation imbalance, identifying correlation-induced and layer-wise imbalances.
• Proposed the Parameter Variation Balancing Framework (PVBF).
• PVBF incorporates parameter correlation computation, an encourage-and-consolidate (E&C) gradient adjustment method, and a dual-layer copy weights with reinit (D-CWR) strategy.

Main Results:

• The proposed Parameter Variation Balancing Framework (PVBF) significantly reduces prediction bias in OCL.
• PVBF achieved up to 47% higher accuracy compared to existing ER-based methods on both short and long task sequences.
• Demonstrated effectiveness in improving OCL performance by addressing parameter update imbalances.

Conclusions:

• Parameter variation imbalance is a critical factor affecting prediction bias in ER-based OCL.
• The PVBF effectively mitigates imbalances, leading to improved OCL performance and accuracy.
• The findings offer a promising direction for developing more robust and adaptive AI systems for continual learning scenarios.