The neural basis of probabilistic computation

Summary

Date: 
December 13, 2016 - 12:00pm - 1:00pm
Location: 
Northwest 243
About the Speaker
Name: 
Wei Ji Ma
Speaker Affiliation: 
NYU

Behavioral models suggest that the brain takes sensory uncertainty into account in decision-making, even as uncertainty varies on a trial-by-trial basis; I call this "probabilistic computation". But is sensory uncertainty a neurally meaningful concept? I will present data from monkey physiology and human fMRI showing that uncertainty can be decoded from visual cortex on a trial-to-trial basis and helps predict behavior, even when conditioning on the stimulus. Next, I will show how simple neural networks can learn to perform probabilistic computation even when trained with errorfeedback only. The representation of uncertainty in the hidden layer of these networks differs by task in a predictable way. Finally, I will caution against confusing probabilistic computation with optimal or Bayesian computation.