Abstract
Cortical activity encodes information about the external world. This cortical activity fluctuates over time. Accordingly, the information encoded by the neural activity can fluctuate over time, with important consequences for behavior. What creates these trial-by-trial changes in the amount of information contained in neural activity? To address this question, it is important to consider that information is predominantly influenced by the tuning of neural populations to features and the noise correlations between different neurons. At the neural level, both theoretical and experimental work has shown that in the limit of many neurons information is bounded primarily by the tuning of the neural populations and a particular form of noise called differential correlations (Moreno-Bote et al., 2014). Whether this combination of noise and tuning similarly constrains the amount of information in trial-by-trial neural population activity as obtained with fMRI is currently unknown. Here, we combine theoretical analysis, simulations, and analysis of fMRI data to assess the impact and structure of noise correlations and tuning in visual cortical activity to address these questions. Information was quantified using a probability distribution extracted from cortical responses to orientation stimuli (visual areas V1-V3). Our analyses show that differential correlations also exist at the level of fMRI voxels. Specifically, we confirm that in the limit of infinite voxels, information is limited by a combination of voxel tuning properties and differential correlations. However, for the number of voxels measured experimentally, information it is still impacted by non-differential forms of correlation. Consequently, it is important to take this form of noise into consideration when quantifying the information contained in fMRI activity patterns. Finally, we show that trial-by-trial fluctuations in information can arise due to stimulus dependent tuning, as well as other modulatory influences on voxels activity such as a change in gain.