Intrinsic plasticity for reservoir learning algorithms

One of the most difficult problems in using dynamic reservoirs like echo state networks for signal processing is the choice of reservoir network parameters like connectivity or spectral radius of the weight matrix. In this article, we investigate the properties of an unsupervised intrinsic plasticity rule for signal specific adaptive shaping of the reservoir, which is local in space and time and aims at maximizing input–to–output information transmission for each neuron. We show that the rule consistently regulates the neurons’ mean outputs and variances and is robust to learning parameter changes. Simulations reveals that this reservoir adaptation robustly enhances online learning of Backpropagation–Decorrelation recurrent learning for a tenth–order nonlinear NARMA benchmark problem.

• Type: Conference
• Year: 2007
• Where: ESANN
• Authors: Marion Wardermann, Jochen J. Steil
• Keywords: Backpropagation–Decorrelation Recurrent Learning, ESANN 2007, Intrinsic Plasticity Rule, Neural Networks, Reservoir Network Parameters