Browsing by Author "Väätäjä, Sara"

This Bachelor’s Thesis covers the IDyOM (Information Dynamics of Music) model, which models the prediction of melody, and is based on statistical learning of melodic regularities. In addition the model is contrasted with the earlier Implication-Realization theory (IR theory) of the prediction of melody. The IR-theory hypothesizes that the prediction of melody is composed of innate bottom-up rules and learned top-down style. First some music theoretical concepts and theories and models of prediction are discussed. Melody is an essential part of music at least in Western culture. Melody can be defined as a sequence of tones. The prediction of melody can be studied with formal models, in this case a cognitive model. Predictive cognitive processes allow an organism or artificial intelligence to make a model of its future state and direct its behaviour accordingly. This Thesis also presents the background theories of statistical learning and information theory. The IDyOM model simulates statistical learning which means finding and learning regularities from input, by using unsupervised learning, in which correct answers are not taught to the model. The model is based on Shannon’s information theory which postulates regularities of discrete sources and stochastic processes. The IDyOM model is multidimensional in two ways: it combines the dimensions of the input, e.g. pitch, into new dimensions, e.g. intervals, and uses two inputs to compare a new melody to a broader corpus. The performance of the IDyOM model is compared to behavioral studies with human subjects and the IR theory. The IDyOM model explains human prediction of melody at least as well as a computational implementation of the IR theory. Statistical learning may be a mechanism for learning melodies.