Random and biological network connectivity for reservoir computing: Random Reservoirs Rule! (at Remembering)

Reservoir computing is a framework where a fixed recurrent neural network (RNN) is used to process input signals and perform computations. Reservoirs are typically randomly initialised, but it is not fully known how connectivity affects performance, and whether particular structures might yield advantages on specific or generic tasks. Simpler topologies often perform equally well as more complex networks on prediction tasks. We check performance differences of reservoirs on four task types using the connectomes of C. elegans and drosophila larval mushroom body in comparison with varying degrees of randomisation.