Self-organizing reservior computing with dynamically regulated cortical neural networks

Abstract

Reservoir computing (RC) is an approach to design, train and analyze recurrent neural networks (RNNs). Traditional RC models use fixed, random topological reservoirs. Very little work has been conducted on adapting the structure of neural reservoirs. In this study, the lamina-specific cortical neural network is employed to construct the reservoir. Numerous experimental data show that cortical neural networks are not silent even without external inputs, but rather maintain a low spontaneous firing activity. This aspect of cortical networks is important for their computational functions, but it is difficult to reproduce this aspect in the models of cortical neural networks due to the inherently unstable property of the low-activity regime. In this paper, we propose a developmental approach to self-organize the structure and synapses of lamina-specific cortical neural networks (CNNs) based reservoir where a gene regulatory network (GRN) is employed. We call this new model as GRN-CNN-RC model. Basically, the weight plasticity, and meta-plasticity of the CNN-based RC will be regulated by the GRN, and the GRN will also be influenced by the activity of the neurons it resides in, in a closed loop. Extensive experimental results on the classical benchmark problems have demonstrated the efficiency and robustness of the proposed model for classification tasks.