Individual Neurons - Computational Explorations in Cognitive Neuroscience - page 24

entist's job that much more difficult, because it quickly
becomes hard to figure out what kinds of input infor-
mation a given neuron is basing its detection decision
on when it is so indirectly related to the sensory inputs.
Regardless of where it gets them, the detector then
performs processing on its inputs. In a smoke detector,
this might involve combining and weighing the inputs
from different types of smoke receptors to provide the
best overall assessment of the likelihood of a fire. In the
detector model of the neuron, the relative contribution
of the different inputs to the overall detection decision is
controlled by weights , which correspond in the neuron
to the relative efficiency with which a synapse commu-
nicates inputs to the neuron (also known as synaptic
efficacies or synaptic strengths ). Thus, some inputs
“weigh” more heavily into the detection decision than
do others. These weights provide the critical param-
eters for specifying what the neuron detects — essen-
tially, neurons can detect patterns of activity over their
inputs, with those input patterns that best fit the pattern
of weights producing the largest detection response.
It does not appear that much neural effort is spent an-
alyzing individual inputs as such, treating them instead
as just a part of the overall input pattern. This is done
by combining or integrating over all the weighted in-
puts to form some sort of aggregate measure of the de-
gree to which the input pattern fits the expected one.
This happens through the electronic properties of the
dendrites, resulting eventually in a membrane poten-
tial (electrical voltage) at the cell body (the central part
of the neuron), which reflects the results of this combi-
nation. Thus, as we will see below, the neuron can be
thought of as a little electronic system.
After integrating its inputs, a detector needs to com-
municate the results of its processing in the form of an
output that informs anyone or anything that is listen-
ing if it has detected what it is looking for. In a smoke
detector, this is the alarm siren or buzzer. Instead of
directly communicating the value of the integrated in-
puts, many detectors have a threshold or criterion that
is applied first. In a smoke detector, the threshold is
there so that you are not constantly bombarded with the
information that there is not sufficient smoke to be con-
cerned about. You only want to hear from a smoke de-
tector when there really is a likely cause for alarm —

output axon

cell body,
membrane
potential

integration

dendrites

inputs

synapses

Detector Neuron

Figure 2.1: The detector model can be used to understand
the function of corresponding neural components.

2.2.1 Understanding the Parts of the Neuron Using

the Detector Model

The neuron, though tiny, is a highly complex biochem-
ical and electrical structure. Any attempt to incorpo-
rate all of its complexity into a cognitive-level sim-
ulation would be impossible using current computa-
tional technology. Thus, simplifications are necessary
on purely practical grounds. Furthermore, these sim-
plifications can help us figure out which aspects of the
neuron's complex biology are most cognitively relevant
— these simplifications constitute an important part of
the scientific endeavor (witness the popularity of the fic-
tional frictionless plane in physics). One of the most
commonly used simplifications of the neuron is the
integrate-and-fire model (Abbott, 1999). As shown
in figure 2.1 and elaborated below, the detector model
maps quite nicely onto the integrate-and-fire model of
the neuron.
First, a detector needs inputs that provide the infor-
mation on which it bases its detection. A neuron re-
ceives its inputs via synapses that typically occur on its
dendrites , which are branching fingers that can extend
for somewhat large distances from the cell body of the
neuron. In the human brain, only a relatively few neu-
rons are directly connected with sensory inputs, with the
rest getting their inputs from earlier stages of process-
ing. The chaining of multiple levels of detectors can
lead to more powerful and efficient detection capabili-
ties than if everything had to work directly off of the raw
sensory inputs. However, this also makes the neurosci-