Reservoir Self-Control for Achiev- ing Invariance Against Slow Input Distortions

A method is presented for immunizing an Echo State Network against slow, task-irrelevant variation in its input. The main idea is to extract some principal components of the smoothed reservoir dynamics, and augment the reservoir with a feedback controller which attempts to pull these components to zero. This leads to a “homeostatic” self-stabilization of the reservoir dynamics. A proof-of-principle case study is presented where the network has to predict a two-mode input signal which is subjected to a large-amplitude, slow variation in offset. With the controller in place, this task is solved with exactly the same quality as in a condition where the input is undistorted. Furthermore, the action of the controller can be “compiled into” the reservoir, by recomputing reservoir weights such that the controlled network dynamics is recovered without the controller switched on. One thus finally obtains a reservoir whose internal dynamics are largely invariant against slow distortion in the input.