Supervised STDP Learning with Weight Decay for Spiking Neural Networks

This research aims to address the challenges of the Diehl & Cook model and develop a new learning method focused on the learning efficiency of spiking neural networks (SNNs). The identified challenges include the task of classification, learning redundancy, and catastrophic forgetting. To overcome these challenges, a novel approach called supervised STDP learning was proposed, leveraging the local learning nature of the STDP learning rule. Additionally, a weight decay mechanism was introduced to assign inhibitory roles to irrelevant weights. Experimental evaluations were conducted using the MNIST dataset, comparing the proposed method with the Diehl & Cook model. The results demonstrated that the proposed method significantly improved accuracy compared to the previous model, and the learning convergence speed was remarkably fast. The experiments also confirmed that supervised STDP learning allows for local learning within synapses, enabling the network to learn independently for each label without interference. As a result, the proposed method showed immunity to catastrophic forgetting. Overall, this research contributes to enhancing the learning efficiency of SNNs and provides insights into overcoming challenges in spiking neural network models.