UWA
Thursday 24 February 2022
Presentation made as part of confirmingPhD candidature.
Reservoir computing (RC) is a framework for supervised machine learning tasks built upon a foundation of dynamical system theory.
The underlying idea is to take some time series and inject it into a high dimensional, fixed recurrent network in order to construct a representation of the data in the reservoir state space.
This state space is rich with information that can then be trained to some desired output or set of labels, and extensive research has shown impressive results for such tasks.
However, in cases where we are not afforded some target output or set of labels, is there still a way in which we can utilise the RC framework?
The ability to use reservoir computing in such an unsupervised would open up new possibilities for the field, particularly for time series analysis.
Here, I’ll introduce reservoir time series analysis as the study of generating and utilising features generated from the reservoir state space to analyse the underlying time series.
I’ll present three ad-hoc feature generating methods which will then be explored for the task of distinguishing signals from systems with varying parameters, presenting a testing framework for determining the capability of methods for this task.
This framework will allow us to highlight the strength and generality of the ad-hoc methods with respect to more well known time series analysis techniques, such as Fourier or recurrence quantification analysis.
Finally, we’ll motivate the use of reservoir time series analysis in application by looking at our ongoing research in applying our proposed features to unsupervised learning tasks, namely one-class classification and concept drift detection.