Document Actions

You are here: Home / Bernstein Conference / 2017 / Satellite Workshops / Neural sampling: computations and experimental predictions (Haefner, Orban)

Neural sampling: computations and experimental predictions (Haefner, Orban)

Abstract

Bayesian theories of perception and decision-making as well as ideal observer analyses have been used for decades to understand sensory processing and decision-making in the brain. Over the past 5-10 years, the algorithm by which those computational goals are achieved (Marr, 1982) has attracted increased attention. In particular, work on the 'neural sampling hypothesis' (Fiser et al. TICS 2010) has resulted in a series of original papers (e.g. Hoyer & Hyvarinnen NIPS 2003, Berkes et al. Science 2011, Buesing et al. 2011, Pecevski et al. 2011, Legenstein & Maass, 2014 (all PLoS CB), Savin & Deneve, 2014, Hennequin et al. 2014, Kappel et al. 2015 (all NIPS), Haefner et al. Neuron 2016, Orban et al. Neuron 2016, Aitchison & Lengyel PLoS CB 2016) that have addressed various computational and empirical aspects of this hypothesis, some of which were discussed in recent reviews (Sanborn & Chater TICS 2016, Gershman & Beck 2016).
This workshop has 3 primary goals:
1) The 'neural sampling hypothesis' represents a large class of theories that differ in the nature of the variables represented (discrete or continuous), in the nature of the representation (membrane potentials or spikes), in the nature of the algorithm (number of concurrent sampling chains or particles, specific sampling scheme, etc.) and in the nature of the underlying probabilistic model. By bringing together 'neural sampling' researchers we hope to make explicit commonalities and differences between the various approaches, discussing their computational pro's and con's and the empirically testable predictions in which they differ.
2) Discussion of the main challenges to 'neural sampling' and how to overcome them.
3) Outreach to and education of the wider audience of the Bernstein Conference

'Neural sampling' connects all 3 levels of Marr's analysis of the brain, linking the computational goal of perception as probabilistic inference to observable neural responses. It is timely since for the first time, a critical mass of related papers has been published warranting a taking stock of existing results and refocusing future research efforts. It parallels related work in cognitive sciences making it a candidate for a fundamental computational theory of cortical function.

Schedule