Individual Neurons - Computational Explorations in Cognitive Neuroscience - page 64

task. However, conspiring against this idea is the huge
amount of biological complexity present in the neuron,
and the vast amount of information processing power
that a neuron could potentially exhibit. For example,
Shepherd and Brayton (1987) and others have argued
that the dendrites of a neuron could potentially perform
complex processing on neural inputs, including various
logical operations (AND, OR, NOT, etc.). Further, a se-
quence of output spikes from a neuron could potentially
convey a huge amount of information by varying the
timing between spikes in systematic ways (c.f., Reike
et al., 1996). Thus, each individual neuron could be
more like a complex CPU instead of a simpler detector.
Although there is clearly room for different perspec-
tives on the relative complexity of neural processing,
our view is motivated by the following considerations:

Each neuron receives as many as 10,000 or even
100,000 inputs from other neurons, and yet it only
sends one output signal. If it is computing detailed
logical operations on these inputs or paying atten-
tion to the detailed spike timing of individual inputs,
the complexity and difficulty of organizing such a
large number of such operations would be tremen-
dous. Further, all this complexity and detail must
somehow be reduced down to a single output signal
in the end, which will provide just one tiny fraction
of the total input to other neurons, so it is not clear
what the point of all this complex processing would
be in the end.

Finally, the bottom line is that we are able to model
a wide range of cognitive phenomena with the sim-
ple detector-style neurons, so any additional complexity
does not appear to be necessary, at least at this point. Learning requires a graded, proportional response to

bootstrap changes, as described in the introduction
and later in chapters 4-6. The more a neuron is
viewed as performing lots of discrete logical oper-
ations, or communicating via precise spike timing,
the more brittle it becomes to the changes induced
by learning, and the less likely there will be a ro-
bust, powerful learning algorithm for such neurons.
Without such a learning mechanism, it becomes over-
whelmingly difficult to organize networks of such
neurons to perform effectively together.

2.9 Self-Regulation: Accommodation and

Hysteresis

[Note: This section is optional, because the mecha-
nisms described herein are applicable to a more lim-
ited range of phenomena, and are not active by default
in most simulations. Readers may wish to come back
to this material later when they find a need for these
mechanisms.]
In addition to the integration of inputs and thresh-
olded communication of outputs, neurons have more
complex activation dynamics that can result in the mod-
ification of responses to subsequent inputs as a function
of prior activation history. These can be thought of as
aformof self-regulation of the neuron's response. Al-
though we think that these dynamics make an important
contribution to behavior, and we incorporate them into
some of our simulations (e.g., sections 3.6.5, 8.5.2, and
8.6.1), most of the time we just ignore them for the sake
of simplicity. We can partially justify this simplification
on the grounds that most simulations are concerned only
with the activation state produced in roughly the first
several hundred milliseconds of processing on a given
input pattern, and these self-regulatory dynamics do not
enter into the picture until after that.

The brain is evidently quite robust to noise and dam-
age. For example, the constant pumping of blood
through the brain results in significant movement of
neural tissue, which undoubtedly introduces some
kinds of noise into processing, and there are many
other known sources of noise due to the unreliability
of various biological mechanisms (e.g., neurotrans-
mitter release). In addition, many drugs have sub-
stantial effects on the detailed firing properties of
individual neurons, but their effect on cognition is
graded, not catastrophic. Further, surprisingly high
levels of diffuse damage (e.g., loss of neurons) can
be sustained before it has any measurable effects on
cognition. As we saw earlier, the rate code is very
robust to noise, while the detailed spike timing is not.