Learning Melodic Expectancy: Musical Predictability in an SRN

Descriptive models of music cognition propose various principles as underlying melodic expectancy, however there is very little discussion regarding the processes involved in the acquisition of melodic expectation. To explore the potential role of learning processes, a simple recurrent network (SRN) was trained on a set of musical sequences to examine the degree to which the principles described by Schellenberg’s (1997) twofactor model might be learned through musical exposure. The principle of pitch proximity, but not pitch reversal constrained the model’s expectations of tones following melodic fragments. Implications for this model in the area of music perception and sequential modeling are discussed, as are potential extensions of this simple system.