A supervised learning algorithm based on spike train inner products for recurrent spiking neural networks

For recurrent spiking neural networks (RSNNs), constructing an efficient supervised learning algorithms is difficult because of their complex recurrent structure and an implicit nonlinear spike firing mechanism. This paper presents a supervised learning algorithm based on spike train inner products in RSNNs. The proposed algorithm transforms the discrete spike train into a continuous function using a special kernel function, and we design the corresponding error function for the backpropagation process. The proposed algorithm is successfully applied to spike train learning and pattern classification problems. The experimental results show that our algorithm has higher accuracy than the algorithm for feedback-based online local learning of weights (FOLLOW). Therefore, it is an effective method to solve the spatio-temporal pattern learning problems.