EnsembleSTDP: Distributed in-situ Spike Timing Dependent Plasticity Learning in Spiking Neural Networks

This paper introduces a novel ensemble Spike Timing Dependent Plasticity (EnsembleSTDP) approach for implementing Spiking Neural Networks (SNNs) with on-chip, online (in-situ) unsupervised learning to accelerate training. EnsembleSTDP trains multiple sub-models concurrently on separate data subsets, followed by a merging and compression process at a centralized node, achieving faster learning. We explore various merging and compression methods by identifying overlapping neurons post-training to reduce model size. Experimental results show that EnsembleSTDP, with two and five remote nodes, accelerates training on the MNIST dataset by approximately 1.98x and 5.02x, respectively, with a 0.41% accuracy gain and a 3.45% accuracy loss relative to the baseline. The source code of EnsembleSTDP learning method is available at https://github.com/klab-aizu/EnsembleSTDP.