Events

Reservoir computing is a powerful tool in learning temporal dynamics and forecasting time series. Although applications often have a stochastic flavour, the existing theory mainly deals with deterministic problems. This poses a significant gap between theory and practice, which we fill in this talk. We do so by establishing a stochastic version of the so-called echo state property, which is the key property needed for successful implementation. We will see that this equates to studying the behaviour of state-space systems as generative models and that the stochastic theory is much richer and much more intricate than its deterministic counterpart.