A STOCHASTIC MODEL USING EASY MEASUREMENTS EXACTLY PREDICTS SPIKE TRAINS AND THE INFORMATION THEY CONTAIN FOR SINGLE NEURONS AND POPULATIONS

Barry J. Richmond

Chief, Section on Neural Coding and Computation

Laboratory of Neuropsychology

National Institute of Mental Health

Bethesda, MD

We have been trying to identify all of the aspects of neuronal responses that carry information, and exclude the aspects that can be legitimately called noise (i.e., aspects that are due to true random variations). We have developed a model that takes simple, traditional measurements, the spike count distribution and the shapes of the spike densities over long periods up to several hundred milliseconds, and, using a random variable, generates spike trains that are indistinguishable from the recorded ones. Applying this result to pairs of neurons recorded from inferior temporal cortex and motor cortex during cognitive tasks (match-to-sample in IT, and instructed movement in M1) shows that the patterns of spikes, including simultaneous ones, are directly predictable from this model. This exact predictability has been seen now in visual areas LGN, V1, and TE, and in M1. Thus, it remains to be shown that there are any neurons for which exact spike timing is not stochastically related to the spike count and response profile over relatively long periods. These traditional measurements, the spike counts, the spike density (and to a small extent the interval distribution) thus appear to constitute a complete description of neuronal spike trains. We have now developed a closed form treatment of this model making it straightforward to both generate simulated trains and to estimate all of the information carried in the trains (including from timing at all resolutions) from straightforward calculations.