Abstract:
A reservoir computer is a type of a recurrent neural network where most of the weights are random, and only a single layer of outer weights is trained by linear regression. Remarkably, this simple training procedure is sufficiently general that reservoir computers are still capable of universal approximation. Consequently, reservoir computing has become an interesting theoretical approach to machine learning problems involving time series, including time series forecasting and classification, as well as value function approximation in reinforcement learning.
This talk will be focused on analyzing a reservoir computer where the input is obtained from a deterministic or stochastic dynamical system, and considering questions like whether it is possible to forecast the trajectory of the system, learn topological invariants, and with what level of uncertainty?