Fig. 8: Demixed PCA unmixes variance associated with trial types (hit, miss, and FA) in two independent components.

A, B MUA activity from the three different trial types (hit trials in blue, miss trials in green, and false alarm trials in red), computed on pre-target (solid lines) or pre-cue (dashed lines) neuronal activities, projected onto the first (A) and second (B) components that maximally explain trial type in one representative session, during the time period of −300 ms to 0 ms locked to the stimulus onset. C, D Boxplot representing the median and the interquartile range of the projected MUA activity onto the first (C) and the second (D) demixed components across sessions (two-sided Wilcoxon signed-rank test *p < 0.001; **p < 0.0001). E Box plot corresponding to the distribution of the cross-validated classification accuracies of linear classifiers given by the two demixed principal components associated with trial types shown in (A) and (B) (two-sided Wilcoxon signed-rank test, ***p = 8e-5). The horizontal dotted line corresponds to the theoretical chance level (33.3%). The Gray boxplot shows the distribution of classification accuracies expected by chance as estimated by 100 iterations of the shuffling procedure (maximal accuracy value obtained across all iterations is considered). Lower and upper box boundaries in boxplots reflect the 90th and 10th percentiles, respectively, the line inside the box reflects the median, and lower and upper error lines show the min and max value of each distribution, respectively.