Neural correlations and predictive coding

The Information Bottleneck is a powerful method in channel coding where input is maximally compressed while maintaining some fixed amount of information about a user-defined variable of interest. It has been applied across several disciplines including neuroscience, where it has been used to show that the retina is capable of near optimal compression of incoming stimuli to maintain as much information as possible about the future of the stimulus. While a great deal of work has examined the pairwise correlation structure of the retina using the inverse Ising model, little has been done to investigate how the constraint of optimal prediction shapes these pairwise correlation structures. Using novel methods in energy-based modeling, we examine the changes in both the stimulus induced and intrinsic pairwise correlations. Across different stimuli with changing time scales, we find signs of adaptation to maintain optimal prediction, as well as separate low dimensional subspaces containing the intrinsic and stimulus correlations. The activity in the stimulus subspace is found to be orthogonal to the activity of the intrinsic subspace, which is a necessary condition for maximal information about the stimulus. We also present evidence that the intrinsic pairwise correlations encode an error signal, sending information about the magnitude of the prediction error downstream.