Reservoir computing
A form of Recurrent neural network. Instead training every weight in the network, a reservoir computer uses rich dynamics generated from a neural reservoir that have recurrent connections (e.g., a random graph). Output neurons are connected to the reservoir and only those weights between the output neurons and reservoir are trained.
- Reservoir Computing Properties of Neural Dynamics in Prefrontal Cortex
- Schrauwen2008computational: This paper provides an explanation on why network structure matters more with binary or spiking neurons. See also Rodriguez2019optimal.
- Gauthier2021next - “next generation” reservoir computing by using the equivalence to Nonlinear vector autoregression.