Does cortical feedback convey learned priors?

Summary

Date: 
November 21, 2017 - 12:00pm
Location: 
Northwest Building, Room 243
About the Speaker
Name: 
Richard Born
Speaker Affiliation: 
HMS, Dept. of Neurobiology

It has been appreciated, at least since the time of Helmholtz (1867), that the retinal image is insufficient for determining what is ‘out there’ in the world. Because of this, the visual system must infer the most likely external cause of its current internal state. To do this optimally, the brain should combine incoming sensory data (likelihood) with stored information (prior probability) to infer the probability of external causes (posterior probability). While both perceptual and neurophysiological studies support a framework of perception as probabilistic inference, strong experimental tests of the framework have been hampered by the fact that we do not understand the nature of the brain’s internal model for general vision. In the ongoing project I will describe, we have attempted to overcome this obstacle by perturbing the internal model through perceptual learning. We combine a theoretical framework, developed by Ralf Haefner and colleagues, which makes testable predictions for the effect of task learning on sensory responses, with an experimental approach involving the manipulation of cortico-cortical feedback during task performance while we record from many sensory neurons simultaneously. Specifically, we monitored neuronal activity in primary visual cortex (V1) while macaque monkeys learned two different versions of an orientation discrimination task, and then learned to switch between task versions on a trial-by-trial basis. Consistent with predictions of the theory, we observed task-related noise correlations in V1 after, but not before, learning. Even when trials of the two different task versions were randomly interleaved, the correlation structure tracked the current task, indicating that the task-version expectations are updated dynamically and reflected in the activity of early sensory neurons. Somewhat surprisingly, however, while a measure of the correlations structure tracked psychophysical performance when the two tasks were interleaved, the same measure was not a good indicator of performance during early phases of task learning. Even more surprisingly, while minimally inactivating feedback from V2 to V1 lowered noise correlations overall, the signature task-relevant structure to the correlations remained intact. I will discuss how future experimental directions can begin to address these curious findings.