Abstract: In recent years, machine learning methods such as deep learning have proven enormously successful for tasks such as image classification, voice recognition, and more. However, these methods require vast amounts of training data and substantial computational resources. In this talk, I will discuss how a reservoir computer (RC) offers advantages for time series prediction for many complex systems applications, ranging from predicting brain dynamics to predicting social dynamics. An RC - a specific architecture of artificial neural network that offers a "universal" dynamical system - draws on its own internal complex dynamics in order to forecast complex systems. Like many other machine learning architectures, RCs provide a knowledge-free approach because they forecast purely from past measurements without any specific knowledge of the system dynamics. By building a scheme that judiciously combines the knowledge-free prediction of the RC with a knowledge-based model, we demonstrate a dramatic improvement in forecasting complex systems. In addition, we show that we can forecast the dynamics of not only individual time series but also of spatial and networked systems by constructing appropriate parallel RC architectures. Finally, we discuss how tuning an RC to operate close to the critical boundary between quiescence and amplification may offer important performance advantages.
Speaker
Michelle GirvanSFI External Professor & Professor of Physics at the University of Maryland