EEG-based affect states classification using Deep Belief Networks

Abstract

Affective states classification has become an important part of the Brain-Computer Interface (HCI) study. In recent years, affective computing systems using physiological signals, such as ECG, GSR and EEG has shown very promising results. However, like many other machine learning studies involving physiological signals, the bottle neck is always around the database acquisition and the annotation process. To investigate potential ways to address this small sample problem, this paper introduces a Deep Belief Networks (DBN) based learning system for the EEG-based affective processing system. Through the greedy-layer pretraining using unlabeled data as well as a supervised fine-tuning process, the DBN-based approaches significantly reduced the number of labeled samples required. The DBN methods also acted as an application specific feature selector, by examining the weight vector between the input feature vector and the first invisible layer, we can gain much needed insights on the spatial or spectral locations of the most discriminating features. In this study, DBNs are trained on the narrow-band spectral features extracted from multichannel EEG recordings. To evaluate the efficacy of the proposed DBN-based learning system, we carried out an subject-independent affective states classification experiments on the DEAP database to classify 2-dimensional affect states. As a baseline to the proposed DBN approach, the same classification problem was also carried out using support vector machines (SVMs) and one-way ANOVA based feature selection process. The classification results shown that the proposed framework using Deep Belief Networks not only provided better classification performance, but also significantly lower the number of labeled data required to train such machine learning systems.