Munn et al., Cell 187, 7303–7313 (2024)
The question
At the level of individual neurons, activity looks sparse and uncorrelated. At the population level, it looks correlated and redundant. Both have empirical support. Munn et al. (2024) argued these aren't contradictions, they're two views of the same multiscale structure. I wanted to understand it well enough to implement it from scratch.
The algorithm
ICG works by repeatedly merging neurons into larger ensembles based on correlation. At each level, compute the full pairwise Pearson correlation matrix, greedily pair each unit with its highest-correlation partner, and sum their activity traces:
r^l_i = r^(l-1)_i + r^(l-1)_jEach iteration doubles the ensemble size K. You go from individual neurons (K=1) to whole-brain population activity across 13+ scales.
What I found
At fine scales, pairwise correlations cluster around zero. As ensemble size grows, the distribution broadens. The same neurons that looked uncorrelated individually start to look coordinated at the population level. The multiscale structure is real and the algorithm surfaces it.