Correlated Neuronal Activity: High- and Low-Level Views - Computational Neuroscience: A Comprehensive Approach - page 340

across conditions, even when the firing rates did not. Thus, correlations seemed to
code whether one or two stimuli were shown [51].
In general, changes in firing rate pose a problem when interpreting variations in
synchrony or correlations, first, because the latter can be caused by the former, and
second, because the impact of a change in correlation upon downstream neurons
becomes uncertain given a simultaneous change in firing rate. When neural activity
is compared in two conditions involving different stimuli, it is likely that the evoked
firing rates from the recorded neurons will change; even the populations that respond
within a given area may be different. This is one of the main factors that muddles
the interpretation of experiments in which correlations have been measured [72].
The most solid paradigms for investigating correlated activity are those in which
variations in correlation are observed without variations in stimulation and without
parallel changes in firing rate, but fulfilling all of these conditions requires clever
experimental design and analysis.
There are many other studies in which correlations have been interpreted as ad-
ditional coding dimensions for building internal representations. The following are
cases in which the confounding factors just mentioned were minimized. Consider
two neurons with overlapping receptive fields, and hence a considerable degree of
synchrony. Analysis of the activity of such visual neurons in the lateral genicu-
late nucleus has shown [23] that significantly more information about the stimulus
(around 20% more) can be extracted from their spike trains if the synchronous spikes
are analyzed separately from the nonsynchronous ones. In a similar vein, recordings
from primary auditory cortex indicate that, when a stimulus is turned on, neurons
respond by changing their firing rates and their correlations [26]. In many cases the
firing rate modulations are transient, so they may disappear if the sound is sustained.
However, the evoked changes in correlation may persist [26]. Thus, the correlation
structure can signal the presence of a stimulus in the absence of changes in firing
rate.
Finally, the antennal lobe of insects is an interesting preparation in which this
problem can be investigated. Spikes in this structure are typically synchronized by
20 Hz oscillations [90]. When these neurons are artificially desynchronized [55], the
specificity of downstream responses is strongly degraded, selectivity for different
odors decreases, and responses to new odors arise, even though this loss of informa-
tion does not occur upstream. Apparently, what happens is that the downstream cells
— Kenyon cells in the mushroom bodies — act as coincidence detectors that detect
synchronized spikes from projection neurons in the antennal lobe. Kenyon cells have
very low firing rates and are highly selective for odors, so in effect they sparsify the
output of the antennal lobe [62]. In addition, disrupting synchrony in this system has
a real impact on behavior: it impairs odor discrimination [79]. This preparation is
also convenient for studying the biophysical mechanisms underlying such oscillatory
processes [9, 10].
These examples show that the neural codes used to represent the physical world
can be made more efficient by taking into account the pairwise interactions between
neural responses. The degree to which this is actually a general strategy used by neu-
rons is uncertain; the key observation is that, under this point of view, correlations