Frequency detectors can be found at almost all major stages in the auditory
system, including the inner ear (Tasaki, 1954), the auditory nerve (Russell &
Sellick, 1977), the cochlear nucleus (Rose, Galambos, & Hughes, 1959), the in-
ferior colliculus (Semple & Aitkin, 1979), the medial geniculate body (Gulick,
Gescheider, & Frisina, 1989) and the auditory cortex (Merzenich, Knight, &
Roth, 1975). In all those structures, neurons seem to be arrangedtonotopically,
that is, systematically in order of characteristic frequency. Although most of the
studies reporting tonotopy have involved animals, recent positron emission
tomography studies have revealed tonotopic frequency tuning in humans at
the cortical level (Lauter, Hersovitch, Formby, & Raichle, 1985). Many of the
representations used in this chapter are tonotopic, although only their tuning,
and not tonotopy per se, is computationally relevant. The networks that oper-
ate on these representations would function equivalently if the neurons were
arranged randomly while preserving their tuning.
It may at first seem odd to think of frequencies as features, because frequency
is a continuous dimension that is infinitely dense, that is, between the lowest
and highest frequencies, we can detect an infinite number of frequencies. Yet
the brain represents this continuum with a finite set of detectors with charac-
teristic frequencies at discrete points. We are unaware of the gaps between the
characteristic frequencies because each frequency detector responds to a broad
band of frequencies around its characteristic frequency, in accord with its tun-
ing curve, and the response band overlaps with those of other neurons (right-
hand panel of figure 19.1). Any given frequency thus activates an entire family
of inner hair cells to various degrees, with the strongest response coming from
the neuron whose characteristic frequency is closest to the sounded frequency.
This form of representation is calledcoarse coding. Coarse coding enables a
perceptual dimension to be denser than the array of neurons used to perceive
it. (Cones in the retina have only three different characteristic wavelengths, yet
we can discriminate hundreds of different colors).
Coarse coding permits the listener to assimilate small tuning differences to
broad musical categories (such as semitones) while also permitting us to detect
fine degrees of mistuning. The former is enabled because the broadening of the
Figure 19.1
Left: Tuning curve. A frequency detector responds most strongly to a particular frequency—
its characteristic frequency (CF)—and less so to frequencies farther away. Right: Coarse coding
achieved by overlapping tuning curves of frequency detectors with different characteristic
frequencies.
456 Jamshed J. Bharucha