Generalization performance (leave-one-participant-out cross-validation) for fitted computational PCA-space (green bars) and image-based (orange bars) models, as well as mean performance for two fixed predictors (final bars in each panel).

<p>The bars provide group-averaged Pearson correlation coefficients, while individual participants are overlaid in gray markers. Filled bars indicate full model fits, while outlined bars indicate model fits excluding measurement-level population averaging. Statistically significant differences between model variants with and without population averaging are illustrated with black connection lines (p<0.05, see <a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1005604#pcbi.1005604.s015" target="_blank">S4 Table</a> for all pairwise comparisons). An estimate of the maximal performance expected given signal-to-noise levels in the sample is illustrated as a shaded noise ceiling (<a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1005604#sec017" target="_blank">Materials and methods</a>). Performance is plotted in separate panels for perceptual judgments (<b>a</b>), early visual cortex (<b>b</b>), and the fusiform face area (<b>c</b>). For other regions of interest, see <a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1005604#pcbi.1005604.s005" target="_blank">S5 Fig</a>.