Is There a Relation Between the Syntax and the Fitness of an Audio Feature?

Feature generation has been proposed recently to generate feature sets automatically, as opposed to human-designed feature sets. This technique has shown promising results in many areas of supervised classification, in particular in the audio domain. However, feature generation is usually performed blindly, with genetic algorithms. As a result search performance is poor, thereby limiting its practical use. We propose a method to increase the search performance of feature generation systems. We focus on analytical features, i.e. features determined by their syntax. Our method consists in first extracting statistical properties of the feature space called spin patterns, by analogy with statistical physics. We show that spin patterns carry information about the topology of the feature space. We exploit these spin patterns to guide a simulated annealing algorithm specifically designed for feature generation. We evaluate our approach on three audio classification problems, and show that it increases performance by an order of magnitude. More generally this work is a first step in using tools from statistical physics for the supervised classification of complex audio signals.