Toward EEG-based Olfactory Sensing through Spatial Temporal Subspace Optimization

Abstract

Recruiting and training sensory panelists for scent product research can be time consuming and costly. Along with the advent of EEG-based brain imaging technique, objective assessment of scent preference is of high interest in a variety of application domains. In this work we explore the EEG-based scent preference identification method. We first designed an effective and accurate data collection procedure. We proposed a machine learning algorithm, Spatial Temporal Subspace Optimization (STSO), for the discriminative subspace learning and classification modeling. A filter bank contains multiple band-pass filters is used to obtain EEG components from specific frequency ranges. Spatial subspace is constructed by exploring discriminative spatial components to enhance the spatial resolution of the EEG. Through the experiment, we confirm that brain signal can be identified in association with responses to pleasant and unpleasant odors, and there is a temporal pattern of such response because the temporal subspace optimization does improve the prediction result. However, event-related potentials were not present in our EEG data, and we have a discussion on the possible causes and implications. Our preliminary result shows that scent can be identified with moderate accuracy.