themselves as particular states of activity in large aggregates or “assemblies” of neurons, with each neuron providing only a coarse coding of the
stimulus in terms of a graded response (Hebb 1949). This is called the
population hypothesis and has received particular attention in recent years
with the emergence of technologies to record and analyze data from
multiple electrodes in behaving animals (e.g., Eichenbaum and Davis 1998;
Nicolelis 2001). Alternatively, phenomena may be encoded in the brain by
the activity of small, possibly redundant populations of relatively specialized neurons. This is called the single–cell hypothesis and is closely tied to
the idea of “feature detectors” (Barlow 1972). The two hypotheses are not
mutually exclusive, and although the early theoretical literature rejected the
single-cell hypothesis (e.g., Marr 1982), in fact both theories may be considered substantially established. The analysis of conspecific vocalizations
has helped in synthesizing a unified perspective.
2.1.1. Single Cells and Distributed Representations
In the extreme case, the single-cell doctrine has been characterized, or
perhaps caricatured, as the “grandmother cell” hypothesis (see Martin 1994;
Barlow 1995). Imagine a cell that responds always and only whenever you
perceive the face of your grandmother and is a requisite for that perception. The grandmother cell concept is defined by necessity and sufficiency
arguments in relation to the percept. However, where these constraints
have been proposed in other contexts of neural coding, such as motor
control, they have generally not proven to be satisfactory criteria
(Kupfermann and Weiss 1978; Eaton 1983). The grandmother cell hypothesis has difficulty addressing the issues of combinatorial explosion, redundancy, and coverage. For instance, consider the case for a theoretical system
containing N neurons. If the response of one (and only one) neuron codes
for one stimulus, the system can only represent N stimuli. If precepts are
represented at high resolution, or in combination, the number of percepts
can easily overwhelm the number of neurons available to code those percepts. Loss of a single neuron would represent loss of the percept. These
facts combined with the uncertain support in the experimental data suggest
that the limits of a single-cell coding scheme are not realized, at least in
vertebrates.
The extreme case of the distributed population-coding hypothesis is
equally implausible. For the theoretical system containing N neurons, a
fully distributed system could represent at least 2
N stimuli (given a binary
response for each cell). Many more percepts could be represented than
there were neurons, and loss of any one neuron or small set of neurons
would result only in gradual degradation of the information represented in
the network. These are desirable properties that superficially mimic biological neural networks, and highly interconnected networks are attractive
to theoreticians because they are amenable to quantitative analysis.
7. Neuroethology of Vocal Communication
331
stimulus in terms of a graded response (Hebb 1949). This is called the
population hypothesis and has received particular attention in recent years
with the emergence of technologies to record and analyze data from
multiple electrodes in behaving animals (e.g., Eichenbaum and Davis 1998;
Nicolelis 2001). Alternatively, phenomena may be encoded in the brain by
the activity of small, possibly redundant populations of relatively specialized neurons. This is called the single–cell hypothesis and is closely tied to
the idea of “feature detectors” (Barlow 1972). The two hypotheses are not
mutually exclusive, and although the early theoretical literature rejected the
single-cell hypothesis (e.g., Marr 1982), in fact both theories may be considered substantially established. The analysis of conspecific vocalizations
has helped in synthesizing a unified perspective.
2.1.1. Single Cells and Distributed Representations
In the extreme case, the single-cell doctrine has been characterized, or
perhaps caricatured, as the “grandmother cell” hypothesis (see Martin 1994;
Barlow 1995). Imagine a cell that responds always and only whenever you
perceive the face of your grandmother and is a requisite for that perception. The grandmother cell concept is defined by necessity and sufficiency
arguments in relation to the percept. However, where these constraints
have been proposed in other contexts of neural coding, such as motor
control, they have generally not proven to be satisfactory criteria
(Kupfermann and Weiss 1978; Eaton 1983). The grandmother cell hypothesis has difficulty addressing the issues of combinatorial explosion, redundancy, and coverage. For instance, consider the case for a theoretical system
containing N neurons. If the response of one (and only one) neuron codes
for one stimulus, the system can only represent N stimuli. If precepts are
represented at high resolution, or in combination, the number of percepts
can easily overwhelm the number of neurons available to code those percepts. Loss of a single neuron would represent loss of the percept. These
facts combined with the uncertain support in the experimental data suggest
that the limits of a single-cell coding scheme are not realized, at least in
vertebrates.
The extreme case of the distributed population-coding hypothesis is
equally implausible. For the theoretical system containing N neurons, a
fully distributed system could represent at least 2
N stimuli (given a binary
response for each cell). Many more percepts could be represented than
there were neurons, and loss of any one neuron or small set of neurons
would result only in gradual degradation of the information represented in
the network. These are desirable properties that superficially mimic biological neural networks, and highly interconnected networks are attractive
to theoreticians because they are amenable to quantitative analysis.
7. Neuroethology of Vocal Communication
331