Neural network-based improvement in class separation of physiological signals for emotion classification

Computer scientists have been slow to become aware of the importance of emotion on human decisions and actions. Recently, however, a considerable amount of research has focused on the utilisation of affective information with the intention of improving both human-machine interaction and artificial humanlike inference models. It has been argued that valuable information could be obtained by analysing the way affective states and environment interact and affect human behaviour. A method to improve pattern recognition among four bodily parameters employed for emotion recognition is presented. The utilisation of autoassociative neural networks has proved to be a valuable mechanism to increase inter-cluster separation related to emotional polarity (positive or negative). It is suggested that the proposed methodology could improve performance in pattern recognition tasks involving physiological signals. Also, by way of grounding the immediate aims of our research, and providing an insight into the direction of our work, we provide a brief overview of an intelligent-dormitory test bed in which affective computing methods was applied and compared to non-affective agents