ABSTRACT
Connectionist models of musical ability such as tonal composition [11, 31, 41], performance [1, 37] and perception [3, 10, 23] entail representation of properties of music including tonal expectancies and notechord-key hierarchies [2, 18]. Many of the models assume, rather than investigate, the representation of tonal music. Given that neural networks comprise a dual mechanism capable of learning representations and of applying the processes which act upon them, networks can, in principle, be used to explore the construction of representations which may mediate cognition. Thus, it is now possible to ask: how are representations of music constructed during learning, and what form do they take? The question is addressed in the present study using a recurrent network to simulate the first stage of music cognition, namely the construction of a representation of the temporal contiguities implicit in a Western tonal composition. We begin with a discussion of the nature of input to the network, followed by a description of the model, and, finally, we examine the time course of learning. The development of representations is traced through an analysis of the prediction performance of the network at various stages of training. The network was exposed to a single composition to explore the veridical expectancies or “instance-based expectations for events that follow in a particular familiar sequence” [3]. The simulation results indicate that learning involves representation of interactions and interdependencies between components of music such as tone, octave, duration, and accent. As the single composition embodied melodic, harmonic, and rhythmic conventions of a particular style of tonal music, the representations constructed by the network also approximated schematic expectancies “for events that typically follow familiar contexts” [3]. More specifically, the distribution of mean activations across the output vector approximated the tonal hierarchy for the key of C major [19, 22], in that the notes of the tonic C chord (C, E, G) were most active and highly stable.
