Fig. 2: Neural compression analysis schematic and results (N = 23).

a Principal component analysis (PCA) was performed on neural patterns evoked for each of n trials within a learning block. The number of principal components (PC) required to explain 90% of the variance (k) was used to calculate a neural compression score (1-k/n). We quantified neural compression as a function of problem complexity and learning block; the interaction of these factors reflects changes in the complexity of neural representations that emerge with learning. b A whole-brain voxel-wise linear mixed effects regression revealed a vmPFC region that showed a significant interaction between learning block and problem complexity. See Supplementary Figure 1 for main effect maps of learning block and problem complexity. The nature of the interaction in the vmPFC region is depicted in the middle panel; points represent compression at the cluster’s peak voxel for each participant and the horizontal lines depict the group average. The right graph plots the results of a Bayesian-estimated linear mixed effects regression of neural compression from the peak voxel of the vmPFC cluster. Posterior distributions of coefficients from the regression model are depicted for the factors of learning block (b), complexity (c), and their interaction (b:c). Shaded regions within the distributions represent 95% high-density intervals. These data and regression results are displayed only to demonstrate the nature of the interaction effect in the vmPFC cluster and do not represent an independent statistical analysis.