Mapping spatio-temporally encoded patterns by reward-modulated STDP in Spiking neurons

Abstract

In this paper, a simple structure of two-layer feed-forward spiking neural network (SNN) is developed which is trained by reward-modulated Spike Timing Dependent Plasticity (STDP). Neurons based on leaky integrate-and-fire (LIF) neuron model are trained to associate input temporal sequences with a desired output spike pattern, both consisting of multiple spikes. A biologically plausible Reward-Modulated STDP learning rule is used so that the network can efficiently converge optimal spike generation. The relative timing of pre- and postsynaptic firings can only modify synaptic weights once the reward has occurred. The history of Hebbian events are stored in the synaptic eligibility traces. STDP process are applied to all synapses with different delays. We experimentally demonstrate a benchmark with spatio-temporally encoded spike pairs. Results demonstrate successful transformations with high accuracy and quick convergence during learning cycles. Therefore, the proposed SNN architecture with modulated STDP can learn how to map temporally encoded spike trains based on Poisson processes in a stable manner.