Mutual Information in High-Dimensional Neural Coding

Abstract

Estimating mutual information in large neural populations is limited by the exponential growth of possible neural response patterns. Here we apply a Monte Carlo framework to make this problem tractable by sampling from a conditionally independent population model fit from single-neuron responses. The estimator accurately recovers analytical ground truth, corrects finite-data bias, and scales to empirical recordings containing 100 to 10,000 neurons from visual cortex, hippocampus, and retina. Across systems, stimulus uncertainty decays exponentially with population size, revealing a simple scaling law for high-dimensional neural codes.