All these behavioral and electro-physiological data suggest that the brains of primates, and doubtless of many other species, harbor mechanisms that approximate Bayesian statistics. At the very least, these mechanisms must represent several probability distributions ; represent and store their a priori ; combine several distributions according to the product rule (or by adding their logarithms) ; and finally, identify their maximum a posteriori. Which neural circuits perform these functions ?
According to Alex Pouget and colleagues, computing on distributions is part of the natural repertoire of any population of neurons whose discharge rates are random according to a certain probability law (Beck et al., 2008; Ma, Beck, Latham & Pouget, 2006). Indeed, in most neurons, the variability of neuronal discharges follows a " de type Poisson " law, belonging to an exponential family of distributions, i.e. the probability of observing a certain number of action potentials in a given time interval follows a curve whose variance is proportional to the mean. Each neuron has a tuningcurve that enables it to discharge in response to certain stimuli (for example, depending on the direction of movement on the retina). Each external stimulus (e.g. the speed of a set of moving dots) is thus represented by a vector of neuronal discharges across a population of neurons. Ma et al (2006) then show that Bayes' rule allows these discharges to be considered as the representation of a probability distribution over the stimulus space. The intensity of the discharges corresponds directly to the width of the probability distribution over the stimuli.
