Echo state networks for feature selection in affective computing

Abstract

The Echo State Networks (ESNs) are dynamical structures designed initially to facilitate learning in Recurrent Neural Networks which are normally applied for time series modeling. In this paper we show that the ESN reservoirs can serve as an effective feature selection procedure that improved the discrimination of human emotion valence from EEG signals, a task that belongs to the research field of affective computing. A number of supervised and unsupervised machine learning techniques provided with the new feature vector extracted from ESN reservoir states were comparatively studied with respect to their discrimination accuracy. This novel application serves as a proof of concept for the possibility of extending the usability of the ESNs in classification or clustering frameworks.