Scaling of information in large sensory populations

How effectively does the brain encode information across large numbers of neurons? Many models predict that shared variability (or, noise correlations) will cause information to saturate for even moderately sized population, although empirical evidence in this regime is severely lacking. We studied this prediction using a novel 3D high-speed in vivo two-photon microscope to record nearly all of the hundreds of neurons in a small volume of the mouse primary visual cortex. We presented full field grating with five closely spaced orientations and measured how encoded information grows with population size. Contrary to numerous predictions, we find that information continues to increase for population sizes of several hundred neurons with little sign of saturation. In addition, a decoder ignoring correlations between neurons can still decode the majority of the information in the population. The growth of information with population size is well described by an equation motivated by models of information limiting correlations [1], I(n) = Ion/ (1+en), with e a consistently low value across numerous anesthetized and awake animals, demonstrating that the magnitude of information-limiting correlations is quite small. Finally, we find the empiric correlation structure is consistent with numerous eigenvectors weakly aligned to the population tuning, f˙(j), which can give rise to similar growth. Our results suggest that sensory neural populations represent information in a truly distributed manner and pooling of neural activity within local circuits may be much more eଏective than previously anticipated. The representation in early sensory areas does not appear to be impaired substantially by shared sensory noise