Biological brains mitigate interference by orthogonalizing neural representations of similar memories, thereby preserving stability across tasks in continual learning. However, most existing continual learning approaches for spiking neural networks (SNNs) adopt randomly initialized classifier heads at each step and optimize them with imbalanced data, which often induces representation drift and undermines model stability. In this work, we revisit the role of the classifier head in the continual learning paradigm and propose a pattern separation learning strategy for expandable SNNs in class-incremental learning (CIL). Specifically, we predefine fixed and mutually orthogonal class centers for each class to replace the conventional learnable classifiers, providing stable optimization targets that prevent feature space conflicts and reduce interference between tasks. Combined with dynamically expandable structures that emulate neurogenesis to enhance plasticity, our approach effectively mitigates catastrophic forgetting while maintaining adaptability to novel tasks. Experimental results show that our PS-SNN achieves an average incremental accuracy of 76.42% on the CIFAR100-B0 benchmark over 10 incremental steps. PS-SNN not only surpasses state-of-the-art SNN-based continual learning algorithms but also matches the performance of DNN-based methods, highlighting the potential of integrating biologically inspired pattern separation into neuromorphic computing systems.
Hu et al. (Mon,) studied this question.