Federated Continual Learning via Neuronal Dynamics for Human Gesture Recognition

Federated learning (FL) frameworks have gained widespread adoption in radar-based human gesture recognition (HGR) owing to their privacy preserving advantages. However, existing FL models are mainly designed for static data distributions and are difficult to adapt to incremental tasks in real-world dynamic learning scenarios due to catastrophic forgetting. In this paper, we propose a federated continual learning framework based on neuronal dynamics (FedNeD), enabling models to accommodate sequential task learning. Inspired by the biological mechanisms of knowledge acquisition and memory consolidation in the brain, the FedNeD develops a spike frequency-based neuron selection strategy. This strategy identifies task-relevant neural subpopulations and dynamically activates them for the current task while suppressing irrelevant neurons to reduce interference, thereby balancing knowledge retention and adaptation between past and new tasks in federated continual learning. Furthermore, our proposed FedNeD introduces a cross-client synaptic regularization strategy to avoid the excessive deviation of local parameters from the global optimum, thus enhancing model generalization in heterogeneous scenarios. To validate the performance of the proposed method, we conduct comparative experiments on multiple datasets and compare it with baseline methods. The results demonstrate that our proposed FedNeD achieves the best performance in balancing new knowledge acquisition with catastrophic forgetting mitigation. In addition, our method reduces communication cost by 20% compared to other baseline methods, thereby enhancing its energy efficiency.