Fig 6. Correlation structure and size of functional groups.

(A) Cumulative distribution function of percent change in test set MSE of LM⁺ and LM^- with respect to the unrestricted LM. The exclusion of edges with positive weights results in a large increase of test set MSE while the exclusion of edges with negative weights results in only minimal increases. These results suggest that edges with positive weights are informative while edges with negative weights are uninformative in generating accurate predictions of single trial responses. (B) Distribution of strong weight magnitudes segregated by positive and negative weights. Strong weights are more likely to be positive rather than negative, thus indicating why positive weighted edges are informative to accurate predictions. (C) Mean sizes of the total functional group and the strong, positive, recurrent subsets against the population size (i.e. the number of imaged neurons).