perturbation-learning

Proposes a perturbation-based learning rule for online self-supervised learning in echo state networks that avoids reservoir-size-dependent variance growth by decomposing the learning cost and perturbing only the input-dependent component.