Self-organizing reservoir computing based on spiking-timing dependent plasticity and intrinsic plasticity mechanisms

Reservoir computing (RC) framework is a type of recurrent neural network (RNN) with the ability of powerful computing performance for input sequences. Unlike traditional RNNs which train the connection weights by complex procedures, RC only uses the linear regression method to train its output weights, which is also much easier to implement. However, academic research of optimization of its structure topologies and learning rules is still insufficient. In this paper, we proposed a novel approach which combines spiking-timing dependent plasticity (STDP) and intrinsic plasticity (IP) for self-organizing spiking reservoir computing. Experiments using the Mackey-Glass time series testing benchmark show that the RC model by our method can converge to a stable state in a short period of time, and is effective terms of the computational accuracy.