My thesis research has broadly focused on how neural systems, at various scales, exploit the structure in the environment to perform information processing tasks efficiently. For the seminar, I will mostly talk about two projects focusing on this idea: i) how the olfactory system harnesses disorder to process natural odours and ii) how humans build and use adaptive auditory priors to optimise perception in noisy, dynamic environments.

In the first part of the talk, I will discuss how the early olfactory pathway exploits the structure in a high-dimensional input space to enable learning flexible associations between odour stimuli and behaviour. In contrast to other sensory pathways, the responses and connectivity in parts of the olfactory pathway seem to lack discernible structure.  Drawing on insights from the mathematics of random projections, I will argue that the diffuse, disordered responses of olfactory receptors to the odours at the first stage provide a good representation of the high-dimensional space of odours. I will then focus on the disorder in connectivity observed in the later stages of the pathway and argue that it serves a different purpose by shaping the receptor representation to efficiently enable learning flexible associations between odours and behaviour.

In the second part, I will talk about a project that combines Bayesian modelling and psychophysics to study human perceptual behaviour in noisy and changing environments. We find that subjects construct and use dynamic priors in a way largely consistent with normative models; however, there are also important individual differences in perceptual behaviour due to differences in the quality of the dynamic priors formed by the subjects. We also show that, on a trial-by-trial basis, pupil fluctuations are predictive of prior usage, beyond what can be inferred from subject responses. This is consistent with the hypothesis that the pupil-linked arousal system acts as a dynamic gain control on the salience of incoming sensory information relative to the priors.

Time permitting, I will also briefly discuss a recent project which aims to connect principled notions of model complexity and extend these notions of complexity to explicitly account for the resolution/scale at which we make measurements.