Reservoir Computing with Both Neuronal Intrinsic Plasticity and Multi-Clustered Structure

Abstract

In the echo state networks, both reservoir states and network structure are essential for the performance of reservoir computing. In neuroscience, it has been confirmed that a single neuron can adaptively change its intrinsic excitability to fit various synaptic inputs. This mechanism is called intrinsic plasticity (IP) mechanism in the literature. This adaptive adjustment of neuronal response to external inputs is believed to maximize input-output mutual information. Meanwhile, the existence of multi-clustered structure with small-world-like property in the brain has been strongly supported by many neurophysiological experiments. Thus, it is advisable to consider both the intrinsic plasticity and multi-clustered structure of a reservoir network, rather than a random network with a non-adaptive reservoir response. In this paper, reservoir models with neuronal intrinsic plasticity and multi-clustered structure are investigated. The effects of two types of IP rules on the performance of several computational tasks have been investigated in detail by combining neuronal IP with multi-clustered reservoir structures. The first type is the Triesch’s IP rule, which drives the output activities of neurons to approximate exponential distributions; another is the Li’s IP rule, which generates a Gaussian distribution of neuronal firing. Results show that both the multi-clustered structures and IP rules can improve the computational accuracy of reservoir computing. However, before the application of the IP rules, the enhancement of computational performance for multi-clustered reservoirs is minor. Both IP rules contribute to improvement of the computational performance, where the Li’s IP rule is more advantageous than the Triesch’s IP. The results indicate that the combination of multi-clustered reservoir structures and IP learning can increase the dynamic diversity of reservoir states, especially for the IP’s learning. The adaptive tuning of reservoir states based on IP improves the dynamic complexity of neuronal activity, which helps train output weights. This biologically inspired reservoir model may give insights for the optimization of reservoir computing.