Reproduce “Similarity-based context aware continual learning for spiking neural networks” on OpenResearch

About this paper

Bing Han, Feifei Zhao, Yang Li, Qingqun Kong, Xianqi Li, Yi Zeng

SCA-SNN is a Spiking Neural Network continual learning algorithm that employs similarity-based context assessment to adaptively reuse and expand neurons. This approach achieves superior accuracy on various benchmarks while significantly reducing energy consumption compared to existing SNN and DNN-based methods, mimicking the brain's context-dependent plasticity.

Key insights

• Mimicking brain's context-dependent plasticity, assessing task similarity allows for adaptive and efficient allocation of neural resources (reuse vs. expansion).
• The "use it or lose it" principle, implemented via gradient-based selective pruning, effectively filters irrelevant knowledge and mitigates catastrophic forgetting in SNNs.
• Linking the number of new neuron expansions and the strictness of pruning directly to task similarity enables a balance between knowledge transfer and avoiding interference, leading to higher efficiency.

Summary by alphaXiv.