Good noise or bad noise? The role of correlated variability in a probabilistic inference framework

Summary

Date: 
March 12, 2014 - 1:00pm
Location: 
NW 243
About the Speaker
Name: 
Ralf Haefner (Born Lab)

The responses of sensory neurons in cortex are variable, and this variability is often correlated [Cohen and Kohn, 2011]. In the first part of my talk I will briefly review how our view of noise correlations has changed from being detrimental [Zohary et al., 1994], to "it depends" [Abbott & Dayan, 1999], to "don't matter" [Shamir and Sompolinsky, 2006; Ecker et al., 2011] to "only 'bad noise' is bad" [Pitkow et al., SfN 2013]. However, it is important to note that the mathematical model underlying this series of works is one of feedforward information processing and ignores the role of feedback influences on sensory responses.

In the second part of my talk I will present my results on the structure of the correlations induced by such top-down connections – assuming that the brain performs probabilistic inference over the task-relevant variables. I will show that if the brain has correctly learnt the task, the structure of task-induced correlations is the same as that of 'bad noise'. However, unlike in the traditional encoding/decoding framework, here their presence reflects the fact that the brain has correctly learnt the task. Furthermore, perceptual learning will increase their amplitude such that stronger correlations of the 'bad noise' structure are reflective of increased top-down task-knowledge (and better performance). Such an increase also entails a steeper relationship between choice probabilities and neurometric performance as has been observed empirically during perceptual learning [Law and Gold, 2008]. Finally, I will show how to infer the internal model used by the brain in a given task from the structure of the noise correlation matrix.