Music Recognition: An Illustrative Application of a Connectionist Model

Abstract

A connectionist model of an atonal discrimination task is reported which illustrates the fundamental principles of artificial neural networks and embodies the assumptions of pattern recognition theory. Musical sequences are defined as patterns consisting of local and global features and it is proposed that recognition of music is achieved by way of processes which extract and differentially weight such features. Musical training serves to refine the feature extraction and weighting processes. As hypothesised, musically trained and untrained listeners were able to discriminate between atonal sequences on the basis of rhythmic and intervallic features although there was no effect of musical training on accuracy and response time measures. Neural network and human data were compared and testable predictions generated by the mechanistic model are provided. The potential contribution of connectionist models to developmental and environmental aspects of music perception and cognition is discussed.