Excitation-Inhibition Balance:
The Search for Functional Principles of Cortical Networks
Neural circuits show a great deal of variability in their
responses, even when identical stimuli are presented. The
source of this variability has been an ongoing question
in studies of neural dynamics. The large magnitude of the
variability is especially puzzling in view of the fact that
individual neurons integrate large numbers of inputs, and
therefore would be expected to average out noise. However,
model neural networks that are constructed with (1) large
numbers of neurons, (2) sparse connections between neurons,
and (3) an approximate balance between excitatory and inhibitory
input can reproduce the high levels of variability seen
in real neural circuits.
We have constructed such models and explored their properties
in great detail using a combination of analytic techniques
and computer simulation. In previous models, noise was always
injected into models from an unknown external source. It
did not arise naturally from the model itself as it does
in this work. The general level of variability, the statistics
of individual response, and the correlations between responses
seen in the model networks match quite well with those measured
in vivo. Furthermore, we have applied these ideas to models
of primary visual cortex and accounted for a number of features
of visual responses seen in recordings from anesthetized
and awake animals. Thus, we feel we have a basic understanding
of the sources and the consequences of high levels of variability
in neural circuits.
