Implementation of exemplar-based learning model for music cognition

"We tend to think of what we `really' know as what we can talk about, and disparage knowledge that we can't verbalize." [Dowling, 1989]

The exemplar-based learning model is proposed here as an alternative approach to modeling many aspects of music cognition. The implementation of this model is based on a k-NN (nearest neighbor) classifier, and on a genetic algorithm which is used for feature weighting.

Although humans are capable of consciously abstracting concepts and deriving rules, there are other cognitive tasks such as music knowledge acquisition that are largely non-verbal and defy generalizations, consequently making the application of traditional rule-based AI models problematic.

In exemplar-based learning, such as the k-NN rule, objects are categorized by their similarity to one or more sets of stored examples, which are represented as weighted feature vectors. Similarity is often defined as the distance between the vectors. In the current implementation, the genetic algorithm is used to find a near-optimal set of weights.

This paradigm, also known as the lazy learning model, is attractive because training is not necessary, learning is extremely fast, algorithms are simple and intuitive, rules are not sought, and learning is incremental. The major drawback has been the high memory requirement, since all examples must be stored, but the recent decrease in memory cost makes this model quite feasible.

Exemplar-based recognition models have been successfully applied in weather prediction, cloud identification, natural language translation, and the acquisition of pronunciation skills. Furthermore, cognitive psychologists have found this model evident in human and animal learning. In music, style recognition, harmonization, expressive performance, instrument recognition, and structural analysis are some of the obvious targets for the deployment of this model.

Use of this model is illustrated with an optical music recognition system and a musical instrument identifier which uses only the steady-state (post-attack) portion of instrumental sounds.