Conference on Cognitive Computational Neuroscience

Trial-to-trial response variability -- generally referred to as ``noise" -- is pervasive in the visual system. Yet it remains unknown why the visual system is noisy and how noise supports or obstructs computation. One intriguing hypothesis is that what appears to be noise may actually encode visual representations used for tasks other than seeing (e.g., deliberation, planning, remembering). This hypothesis predicts that noise correlations and signal correlations should have the same structure. To test this prediction, we used the massive 7T fMRI Natural Scenes Dataset to characterize signal and noise in multiple areas along the visual hierarchy. We find that voxel-to-voxel correlation structure in noise is well aligned with voxel-to-voxel correlation structure in signals evoked by natural scenes. This finding generalizes to correlation structure observed in spontaneous activity during rest (viewing a blank screen). Importantly, we find that noise correlations do not match the correlation structure in signals evoked by artificial stimuli. These results suggest that noise in the visual system encodes naturalistic visual representations that are not directly related to the current visual input. We speculate that these putative representations hidden in the noise are useful for solving cognitive tasks.