A Computational Model of Me on Multiple Representations Maps

Perceiving similarity relationships in melodies is a fundamental musical process effortlessly performed by the cognitive system of listeners. For this reason, computationally heavy methods such as string-matching algorithms may not be plausible as perceptually oriented computational models of this process. Instead of direct comparison of musical events, similarity can be thought of as a higher order emergent property that has been abstracted by learning, as a result of which similarity relations are based on more compressed representations, such as prototypes. A computational model of melodic similarity is proposed that is based on multiple representations and feature maps formed by unsupervised learning algorithms (self-organizing maps, SOM). Several commonly used components of melodic similarity, including melodic contour and distributions of tones, intervals and durations, are represented as vectors in multidimensional spaces and similarity of these features as distance in metric space. The SOMs yield clustered projections of these features, thus providing reduced representations in terms of both dimensionality and number of feature vectors. The SOMs can be used as a general tool for examining similarity relationships of the features. The current work makes the following contributions. First, several melodic features are extracted in parallel, each feature being represented in a distributed fashion. This corresponds to current understanding of the functioning of the perceptual system (e.g. multiple representations in the visual domain). Second, the model simulates the learning of melodic schemata through self-organization. Third, the model can be applied to musical data mining. This talk will demonstrate how the proposed computational model of melodic similarity is constructed and how it can be used in musical data mining. Representations of melodic features and similarity relationships generated by the model are discussed in terms of perceptual relevance.