Correlated Neuronal Activity: High- and Low-Level Views - Computational Neuroscience: A Comprehensive Approach - page 338

sense, correlation might appear as a trivial phenomenon. For instance, if one looks at
day-long activity, practically the whole cerebral cortex fires in a correlated manner,
because of the sleep-wake cycle. Similarly, whenever an object appears within the
visual field, many neurons in visual cortex are expected to respond throughout the
same time interval. Clearly, such correlations are to be expected. However, as the
observation time window becomes smaller, explaining the presence of correlations
becomes more difficult and, at the same time, potentially much more useful. Sup-
pose the activity of two visual neurons is monitored during presentation of a visual
stimulus, after its onset. Suppose also that within a short time window of, say, a
few hundred milliseconds, spikes from the two neurons tend to appear at the same
time. Why is this? Neither the sensory information nor the state of the subject are
changing in an appreciable way, so the correlation must reflect something about the
internal dynamics of the local circuitry or its connectivity. This is where correlations
become interesting.
Thus, correlations at relatively short timescales become useful probes for under-
standing what neural circuits do, and how they do it. This is what this chapter is
about. This analysis goes down to the one millisecond limit (or even further), where
correlation changes name and becomes synchrony. Even at this point, the signifi-
cance of correlated activity cannot be taken for granted. Some amount of synchrony
is practically always to be expected simply because cortical neurons are highly inter-
connected [14, 91]. The question is not just whether there is any correlated activity
at all, but whether timing is an issue and correlations make any difference. In other
words, given the function of a particular microcircuit or cortical area, if the system
were able to control the level and timescale of correlated activity, what would the
optimal values be? For example, in a primary sensory area, stimulus representation
is of paramount importance, so maybe measuring an excess of coincident spikes in
this case is not an accident, but a consequence of the algorithm that local circuits use
to encode stimulus features. This is just an example; the broader question is whether
neurons exploit the precise coincidence of spikes for specific functions. There are
several theoretical proposals that revolve around this concept; we discuss some of
them below.
Asking about the functional implications of correlated activity is one way to at-
tack the problem; this is a top-down approach. Another alternative is to take a
bottom-up view and investigate the biophysical processes related to correlated fir-
ing. These come in two flavors, mechanisms by which correlations are generated,
and mechanisms by which a postsynaptic neuron is sensitive to correlated input. In
this case valuable information can be obtained about possible correlation patterns
and timescales, and in general about the dynamics of correlated activity.
This approach is also important because it sheds some light on a fundamental
question: how does a single cortical neuron respond under realistic stimulation con-
ditions? The reason this is a problem is the interaction between spike-generating
mechanisms, which are inherently nonlinear, and the input that drives the neuron,
which typically has a complicated temporal structure. The major obstacle is not the
accuracy of the single-neuron description; in fact, classic conductance-based models
[24] like the Hodgkin-Huxley model [48] are, if anything, too detailed. The larger