Adaptive Meta-Continual Learning with Dynamic Stability-Plasticity Balancing

Building on Kang & Choi's (2024) recall-oriented framework that separates stability and plasticity mechanisms, this research challenges the assumption that fixed architectural separation is optimal. Instead, we propose using meta-learning to dynamically search for the optimal stability-plasticity configuration for each incoming task. The system would meta-learn a controller that adjusts: (1) the ratio between generative recall and inference network usage, (2) parameter sharing strategies between networks, and (3) task-specific regularization strengths. This addresses the conflict in current literature where some approaches (like Mishra et al.'s meta-CL methods) favor algorithmic solutions while others (like Kang & Choi) prefer architectural separation. By making the balance adaptive, we could achieve better performance across heterogeneous task sequences while maintaining computational efficiency—something Zou & Lin's (2022) Taylor expansion methods prioritize but without architectural flexibility.

References:

1. Recall-Oriented Continual Learning with Generative Adversarial Meta-Model. Han-Eol Kang, Dong-Wan Choi (2024). AAAI Conference on Artificial Intelligence.
2. Meta-Learning Based Continual Learning Methods for Fine-grained Fruit Quality Classification. Aayush Mishra, Prathamesh Gadekar, Parijat Deshpande, Manasi S. Patwardhan (2025). 2025 IEEE Applied Sensing Conference (APSCON).
3. Efficient Meta-Learning for Continual Learning with Taylor Expansion Approximation. Xiaohan Zou, Tong Lin (2022). IEEE International Joint Conference on Neural Network.