Abstract
During walking and running, animals display rich and coordinated motor patterns that are generated and controlled within the central nervous system. Previous computational and experimental results suggest that the balance between excitation and inhibition in neural circuits may be critical for generating such structured motor patterns. In this paper, we explore the influence of this balance on the ability of a reservoir computing artificial neural network to learn human locomotor patterns, using mean-field theory and simulations. We created networks with varying neuron numbers, connection percentages and connection strengths for the excitatory and inhibitory neuron populations, and introduced the anatomical imbalance that quantifies the overall effect of the two populations. We trained the networks to reproduce muscle activation patterns derived from human recordings and evaluated their performance. Our results indicate that network dynamics and performance depend critically on the anatomical imbalance in the network. Excitation-dominated networks lead to saturated firing rates, thereby reducing the firing rate heterogeneity and leading to muscle coactivation and inflexible motor patterns. Inhibition-dominated networks, on the other hand, perform well, displaying balanced input to the neurons and sufficient heterogeneity in the neuron firing rate patterns. This suggests that motor pattern generation may be robust to increased inhibition but not increased excitation in neural networks.