A Homeostasis based Enhanced Training Method in Spiking Neural Networks for Pattern Recognition

Spike Timing Dependent Plasticity (STDP) is a commonly used algorithm structure for unsupervised training of Spiking Neural Networks (SNNs). In the training process, appropriately reducing the update frequency of synaptic weights can enhance the training effect of the SNNs, especially when the training samples are insufficient. This paper proposes a synaptic weight update frequency adjustment mechanism based on homeostasis, which effectively improves the recognition accuracy of SNNs trained for MNIST classification tasks. Besides, the number of weight updates has been reduced by more than half, thus improving energy efficiency and facilitating implementation in hardware.