Unsupervised Spectral Clustering of Music-Related Brain Activity

The recent advancements in Music Information Retrieval are now giving birth to new exciting fields, one of which is concerned with understanding the relationship existing between brain activity and the music stimuli evoking it. Thus, Music Imagery Information Retrieval (MIIR) has emerged with its goal being to bridge the gap existing between encephalographic responses and the respective music signal. This paper employs the OpenMIIR dataset which includes synchronized recordings of brain activity and music signals, thus facilitating MIIR research. Three tasks have been defined, i.e. stimuli identification, group and meter classification, which examine the problem from different viewpoints. After extracting parameters of linear time-invariant models elaborating on electroencephalographic responses, we demonstrate a suitably-designed unsupervised spectral clustering scheme. Such a scheme highlights the connection existing between responses and the audio structure of the music classes corresponding to the three tasks. We show that there is a strong connection w.r.t stimuli identification and meter classification tasks; however that is not true for the group classification case.