Effects of Tuning Sharpness on Tone Categorization by Self-Organizing Neural Networks

Abstract

Because humans can categorize tones either in terms of pitch height or pitch class (chroma), I have explored the potential of simple self-organizing networks to demonstrate these two different capacities, and have uncovered a mechanism which can subserve both. A self-organizing neural network architecture was used; during training, the network learns to represent internally the co-existence of stimulus features (here, harmonic components). Two sets of simulations were completed, identical in all respects except for the tuning sharpness used at the input layer. In the broad tuning condition, the network categorized sounds together exclusively on the basis of their pitch similarity. Here, two tones with nearby fundamental frequencies (FOs) activate many of the same input units, due to the overlap between adjacent tuning curves. Thissimilarity in activation decreases with increasing distance between the FOs. In the narrow tuning condition, a different result was found: many tones were categorized together on the basisof chroma similarity, as opposed to pitch similarity. In this case, two tones with nearby FOs do not activate many of the same input units, since the receptive fields of adjacent units do not overlap as much. However, two tones an octave apart still activate many of the same input units, since half of the upper harmonic components of the lower tone are also present in the upper tone. This results in categorization by chroma.