Abstract

Olfactory system faces similar problems to that of vision or audition, whereby a mixture of sources (visual objects, voices or odours) causes a widespread activation of olfactory receptors. To infer what odours are impinging on the receptors is an over complete task, since there are many more odours than receptor types. The problem is exacerbated by the fact that odours rarely occur in solitude, therefore the system needs to infer at least the major components of olfactory mixtures. To this end, one also needs to make a good guess on the (relative) concentrations of the components. In my talk, I will concentrate on an approximate scheme that follows this logic: our algorithm combines a Gibbs sampler that makes guesses on odour components of the scene, and a Langevin sampler that estimates the posterior over concentrations in agreement with the current hypothesis (on which odours are being present). We show how to avoid the problem of synchronous updates in a network of neurons, whose computations naturally run in parallel. Finally, the plausibility of our algorithm will be discussed and compared to an alternative inference scheme, such as variational Bayes (Beck and all, 2012).