Biomedical Engineering Reference
In-Depth Information
accounted for by components must always be considered. With
large data sets, it is tempting to consider smaller components (i.e.,
∼
5). However, only a few neurons usually contribute to such
components and we have found that such components are sel-
dom informative about behavior. Third, caution must be applied
in interpreting the components, in that the activity of a small sub-
set of neurons at any given moment might drive a given com-
ponent. To determine the robustness of the components, one
should examine the scores for each component and check them
against raw plots of neuronal activity, especially for neurons that
have large weights on the components. In addition, we recom-
mend that cross-validation techniques be used to assess the overall
strength of components. Fourth, components themselves may not
be truly orthogonal, and methods for rotating the components
such as varimax or ICA, might be useful for interpretation
(22)
.
These concerns notwithstanding, this section illustrates how
PCA can readily be used to identify common response proper-
ties among a group of simultaneously recorded neurons. Fur-
thermore, these common response properties might correspond
to functional groupings among neurons. While identifying pat-
terns of responses, PCA cannot comment on specific interactions
between neurons. To make such inferences, we move to other
techniques.
5.
Cross-Correlation
In neuroscience, a common and readily interpretable measure
of functional interactions between neurons is cross-correlation
(23, 24)
. This method has been used in many neural systems to
make powerful inferences about functional interactions as well as
functional anatomy. Cross-correlation is designed to make infer-
ences about synaptic connectivity between neurons (
Fig. 7.4A
).
In investigating synaptic connectivity, one should observe nar-
row peaks (a few ms) in the latency of spikes of one neuron
with respect to another. If such a peak is observed, then one
can potentially make inferences about the number of synapses
between each neuron. In certain, highly anatomically connected
systems that are guided by topography (such as the auditory sys-
tem), cross-correlation may provide detailed information about
anatomical connectivity in lieu of intracellular recordings. Cross-
correlation can also be applied to make inferences about local
circuits
(25)
. Such inferences must be made with great care and
require having well sorted single units. In extracellular recordings,
waveforms recorded on the same wire may interact, diminishing
the ability to spike sort. Furthermore, the recording system must
sample at high rates to ensure waveforms do not collide. For these
