Exploring the role of heterogeneity in reservoir computing dynamics
Date:
Oral presentation about heterogeneity in reservoir computing.
Abstract: Reservoir computers (RCs) are a type of neural network that excel at forecasting chaotic dynamical systems. However, principles for tuning their hyperparameters remain unclear, particularly regarding the impact of heterogeneity in individual neuron properties, such as degree (number of connections). A simple one-dimensional mean-field approximation to RC dynamics offers a limited but useful description of average node activation and oscillatory onset, though it applies only to homogeneous-degree RCs. Here, we refine this approximation to explicitly account for degree heterogeneity, enabling more accurate predictions for Erdós–Rényi-like and power-law degree distributions. Our analysis shows that heterogeneity shifts the region of hyperparameter space where non-saturated neuron activation occurs, as well as regions of good task performance, guiding further investigation into hyperparameter tuning in heterogeneous reservoirs.