To explore how information is encoded by the patterns of action
potentials across entire populations of neurons, I collaborated
with Dave Warland and Markus
Meister at Harvard to analyze spike trains recorded simultaneously from many (10-50)
individual ganglion cells of the tiger salamander retina during visual
stimulation. We compared the optimal linear filter reconstruction
to that obtained by training artificial neural networks to decode
spike trains. The input signal to the retina in this case was a
full-field illumination white noise flicker. The retina encodes this
signal using many nonlinear, analog processing stages, but we found
that a linear decoder could extract as much information about the information in
the final spike representation as could the very non-linear neural network.
Eventually it must be possible to decode at least the information necessary to account for the limits of visual discrimination by the animal. Some previous measurements of the visual behavior of salamanders are of interest in this context.