Supervised and unsupervised neural net learning algorithms for modeling music perception

Abstract

Neural net learning algorithms permit the development of physiologically plausible computational models of learning in the auditory domain. Supervised learning algorithms are applicable when the acoustic environment provides target vectors that can guide learning. Musical sequences can be learned in this manner if each successive element in a sequence trains the expectancies generated by an accumulated representation of the sequence thus far [P. M. Todd, Comput. Mus. J. 13, 27–43 (1989); J. J. Bharucha and P. M. Todd. Comput. Mus.] J. 13, 44–53 (1989)]. Even when target vectors are absent, learning of auditory patterns can be accomplished via unsupervised learning algorithms. Both forms of learning seem to be necessary in order to account for a range of phenomena in music perception, including hierarchical and sequential representation of both specific musical patterns as well as abstract patterns that typify cultural regularities. [Work supported by NSF.]