STAER: A Temporal Aligned Rehearsal Spiking Neural Network for Continual Learning

Continual Learning (CL) addresses the problem of learning from sequential data streams, where a model must acquire new knowledge without degrading what it has previously learned. In this setting, one of the main challenges is catastrophic forgetting, namely the tendency of neural networks to overwrite past representations when trained on new tasks. While replay-based strategies have proven effective in conventional Artificial Neural Networks, their application to Spiking Neural Networks (SNNs) introduces additional difficulties, since information in these models is encoded not only in activation values but also in the temporal dynamics of neural responses. This thesis proposes STAER (Spiking Temporal Alignment with Experience Replay), a continual learning method for SNNs that combines experience replay with temporal alignment of logit trajectories. The central idea is that preserving past knowledge in a spiking model should not be limited to maintaining correct final predictions, but should also enforce temporal consistency in the evolution of the response. To this end, the method stores multi-scale temporal traces in the replay buffer and introduces an alignment objective based on the Soft-DTW divergence, applied to compressed, original, and expanded versions of the model response. The experimental evaluation, conducted on Sequential-MNIST and Sequential-CIFAR10 under both Task-Incremental and Class-Incremental protocols, shows that STAER consistently improves over the main replay-based spiking baselines, reducing forgetting and increasing final accuracy, especially in the more challenging Class-Incremental setting. Overall, the results suggest that explicitly modeling the temporal dimension is a key ingredient for making SNNs more effective and competitive in continual learning.