Churchland lab

 By S. TANABE, A. ZANDVAKILI, A. KOHN;  Neurosci., Albert Einstein Col. of Med., Bronx, NY

This group tackled a long-standing question in the field of decision-making: how do you tell the degree to which a single neuron “weighs in” on a behavioral choice? The question is important, but hard to get at via traditional single cell recording methods, especially for a fine discrimination task as the authors used here. 

To get around this problem, this group recorded from 15-30 neurons in V4 while well-trained subjects made decisions about the orientation of a stimulus. As predicted, they didn’t find a strong relationship between the firing rates of single neurons and the subjects’ choices, but things changed dramatically when they looked at the population level and used a linear classifier. What might account for this? The authors argue that in high dimensional state spaces, the decision axis and variability axis are aligned. This means that even if a given neuron is a key player in a decision, the relationship between its firing and the animal’s choice might be weak.  

At the end of the talk, the authors suggested a “trick” for evaluating the degree to which decision and variability axes are aligned: they shuffled the trials and found that sometimes the classifier did better! The effect of trial shuffling on the classifier, they argue, offers insight into the weighting profile of the neurons. I haven’t heard of anyone taking this approach before- it will be interesting to test on other datasets, especially those on coarse discrimination tasks where the prediction differs.