Advances in Multiple Viewpoint Systems and Applications in Modelling Higher Order Musical Structure
MetadataShow full item record
Statistical approaches are capable of underpinning strong models of musical structure, perception, and cognition. Multiple viewpoint systems are probabilistic models of sequential prediction that aim to capture the multidimensional aspects of a symbolic domain with predictions from multiple finite-context models combined in an information theoretically informed way. Information theory provides an important grounding for such models. In computational terms, information content is an empirical measure of compressibility for model evaluation, and entropy a powerful weighting system for combining predictions from multiple models. In perceptual terms, clear parallels can be drawn between information content and surprise, and entropy and certainty. In cognitive terms information theory underpins explanatory models of both musical representation and expectation. The thesis makes two broad contributions to the field of statistical modelling of music cognition: firstly, advancing the general understanding of multiple viewpoint systems, and, secondly, developing bottom-up, statistical learning methods capable of capturing higher order structure. In the first category, novel methods for predicting multiple basic attributes are empirically tested, significantly outperforming established methods, and refuting the assumption found in the literature that basic attributes are statistically independent from one another. Additionally, novel techniques for improving the prediction of derived viewpoints (viewpoints that abstract information away from whatever musical surface is under consideration) are introduced and analysed, and their relation with cognitive representations explored. Finally, the performance and suitability of an established algorithm that automatically constructs locally optimal multiple viewpoint systems is tested. In the second category, the current research brings together a number of existing statistical methods for segmentation and modelling musical surfaces with the aim of representing higher-order structure. A comprehensive review and empirical evaluation of these information theoretic segmentation methods is presented. Methods for labelling higher order segments, akin to layers of abstraction in a representation, are empirically evaluated and the cognitive implications explored. The architecture and performance of the models are assessed from cognitive and musicological perspectives.
- Theses