Chapter 19

Neural Nets, Temporal Composites, and Tonality

Jamshed J. Bharucha

In this chapter, I outline a framework in which aspects of cognition can be un-
derstood as the result of the neural association of patterns. This approach to
understanding music cognition originates with Pitts and McCulloch (1947) and
Deutsch (1969). Subsequent advances (e.g., Grossberg, 1970, 1972, 1976; Rumel-
hart and McClelland, 1986) enable us to understand how these neural associa-
tions can be learned. Models based on these mechanisms are called neural net
models (also connectionist models or parallel distributed models).
Neural net models have a number of properties that recommend them as
models of music cognition. First, they can account for how we learn musical
patterns through exposure and how this acculturation influences our sub-
sequent perception of music. Second, their assumptions are either known or
plausible principles of neuroscience. Third, they shed light on the observation
(Terhardt, 1974) that aspects of pitch and harmony involve the mental com-
pletion (or Gestalt perception) of patterns. Fourth, they are capable of recog-
nizing varying shades of similarity and are therefore well suited to modeling
similarity-based accounts (e.g., Krumhansl, 1990) of tonality or modality. Fin-
ally, they can discover regularities in musical styles that may elude formal
music-theoretic analysis (Gjerdingen, 1990).
Section I of this chapter deals with neural representation, and Section II deals
with neural association and learning.

I. Neural Representation

A. Frequency Tuning of Neurons
Many neurons, particularly sensory neurons, are highly selective in their re-
sponse. For example, there are neurons in the auditory system that respond
selectively to specific bands of frequencies. Within this band, there is usually a
frequency to which the neuron responds maximally (called thecharacteristic
frequency). For the purpose of the present analysis, the signal to which a neuron
responds maximally may be called afeature, and the neuron itself may be called
afeature detector. A neuron that has a characteristic frequency may be called a
frequency detector. A feature detector responds progressively less strongly to
signals that are increasingly dissimilar. This relationship is given by itstuning
curve. The left-hand panel of figure 19.1 shows a schematic tuning curve of a
frequency detector.

From chapter 11 inThe Psychology of Music,2ded.,ed.D.Deutsch(SanDiego,CA:AcademicPress,
1999), 413–440. Reprinted with permission.