The trial-to-trial response variability that is shared between pairs of neurons (termed spike count correlations1 or rSC) has been the subject of many recent studies largely because it might limit the amount of information that can be encoded by neuronal populations. Spike count correlations are flexible and change depending on task demands2-7. However, the relationship between correlated variability and information coding is a matter of current debate2-14. This debate has been difficult to resolve because testing the theoretical predictions would require simultaneous recordings from an experimentally unfeasible number of neurons. We hypothesized that if correlated variability limits population coding, then spike count correlations in visual cortex should a) covary with subjects’ performance on visually guided tasks and b) lie along the dimensions in neuronal population space that contain information that is used to guide behavior. We focused on two processes that are known to improve visual performance: visual attention, which allows observers to focus on important parts of a visual scene15-17, and perceptual learning, which slowly improves observers’ ability to discriminate specific, well-practiced stimuli18-20. Both attention and learning improve performance on visually guided tasks, but the two processes operate on very different timescales and are typically studied using different perceptual tasks. Here, by manipulating attention and learning in the same task, subjects, trials, and neuronal populations, we show that there is a single, robust relationship between correlated variability in populations of visual neurons and performance on a change-detection task. We also propose an explanation for the mystery of how correlated variability might affect performance: it is oriented along the dimensions of population space used by the animal to make perceptual decisions. Our results suggest that attention and learning affect the same aspects of the neuronal population activity in visual cortex, which may be responsible for learning- and attention-related improvements in behavioral performance. More generally, our study provides a framework for leveraging the activity of simultaneously recorded populations of neurons, cognitive factors, and perceptual decisions to understand the neuronal underpinnings of behavior.
Optimal Tuning Widths in Population Coding of Periodic Variables
