A Supervised Learning Algorithm for Recurrent Spiking Neural Networks Based on BP-STDP

Spiking neural network model encodes information with precisely timed spike train, it is very suitable to process complex spatiotemporal patterns. Recurrent spiking neural network has more complex dynamics characteristics because of feedback connections, which makes it difficult to design efficient learning algorithms. This paper proposes a supervised learning algorithm to train recurrent spiking neural networks. By mapping the integrate-and-fire neuron model to the rectified linear unit activation function, the learning rule is induced using error backpropagation and spike-timing dependent plasticity mechanism. The results of spike train learning task and non-linear pattern classification show that the algorithm is effective to learn spatiotemporal patterns. In addition, the influences of different parameters on learning performance are analyzed.