Extended Data Fig. 5: Alternative implementation of the CDM task.

A network trained with different hyperparameters offers an example of an alternative solution for the CDM task, using 3 effective populations and a fourth one accounting for neurons that are not involved in the task (see Supplementary Text 3)). (a) Left: for each number of subpopulations, a boxplot shows the performance of 10 networks with connectivity resampled from a Gaussian Mixture Model (GMM) fitted to the trained network (line: median, box: quartiles, whiskers: range, in the limit of median ± 1.5 interquartile range, points: outliers). Right: for the GMM with four subpopulations, size of each component found by the clustering procedure. (b) Four 2d projections of the 7-dimensional connectivity space. (c) Upper-right triangle of the empirical covariance matrices for each of the four populations. (d) Illustration of the mechanism used by the network at the level of latent dynamics. Populations 2 − 4 control one effective coupling each, indicated by the matching color. (e) Psychometric matrices similar to those shown in Fig. 4 after inactivation of each subpopulation. (f) Violin plots showing the gain distributions of neurons in each of the four subpopulations in each context.