Increased robustness and intermittent dynamics in structured Reservoir Networks with feedback

Recent studies using feedforward Echo State Networks (ESN) demonstrate that reservoir stability can be strongly affected by reservoir substructures, such as clusters. Here, we evaluate the impact of including feedback on clustered ESNs and assert that certain cluster configurations extend the permissible range of spectral radius values. We also report a new class of reservoir activity: intermittent dynamics, characterized by variable periods of chaotic activity before returning to quiescent behaviour. Using a non-linear benchmark data set, we establish clustered ESNs have comparable performance against conventional ESNs, but importantly display an increased tolerance and robustness to spectral radius and input choice.