How to Recognize Emotions Without Signal Processing: An Application of Convolutional Neural Network to Physiological Signals

Abstract To recognize emotions from physiological signals, several steps are required: collection of data during emotion elicitation, signal processing of physiological data, and machine learning. Although this processing chain has shown satisfactory results in prior studies, transferring this solution outside the laboratory context is complex. With the development of wearable devices including physiological sensors, it becomes possible to consider consumer applications. But, to be effective, it is necessary to develop a recognition system able to work potentially in real time on mobile devices. This chapter presents recognition systems based on convolutional neural networks which can be implemented in mobile devices. This solution offers the major advantage of eliminating the signal-processing step. Moreover, a model which is able to predict continuous values of emotion (i.e., valence, arousal, and dominance) was built. The experiments showed an accuracy of 0.76 on these three dimensions.