Abstract

In many situations humans and animals seem to use uncertainty information to guide close to optimal behaviour. The neural underpinnings of such probabilistic computation remain highly controversial (Fiser et al, 2010; Pouget et al, 2013). Contrary to the traditional view of probabilistic population codes (PPCs) and sampling-based representations as mutually exclusive alternatives, here we argue that both are useful data structures, needed at different computational stages for (approximately) optimal decision making. First, sampling allows neural circuits to represent the joint statistics over many features and to marginalise out nuisance variables (Savin and Deneve, 2014). Second, PPCs enable a compact, quasi-instantaneous representation of the marginals which can be easily combined with cost-related information to yield approximatively optimal decisions.

We propose a model in which the two are linked via a probabilistic form of evidence integration akin to that in (Boerlin et al, 2011), with the ‘evidence’ given by samples from the target distribution. We use the same spike-based code (Boerlin, 2011) for representing both the sampling, and the evidence-integration modules at the level of neural activity. The core idea is that the circuit evolves through recurrent dynamics such that the relevant signal (either samples from a target distribution, or a PPC -like representation of any of its marginals) can be linearly decoded from neural responses. This shared encoding means that many classic measures of single cell responses are consistent with experimental data and, moreover, are preserved across processing stages. Thus the model provides a computationally well-justified reconciliation between competing probabilistic neural codes and points to population-level analyses as tools for validating probabilistic computation experimentally.