Reliable decoding of task variables across trial duration from trial-by-trial population activity.

(a) Mean population activity trajectories for an example session, projected onto the first 3 principal components, separate over the course of the trial by stimulus, choice side, and outcome. (b) Decoding accuracy in sliding 100 ms bins over the course of a trial for features of the previous trial (choice side, outcome) and of the current trial (stimulus category, choice side, and outcome). Thin lines correspond to individual sessions, while bold lines denote the mean across sessions. Decoding accuracy is calculated as proportion of test set classified correctly from activity at a given time point, and below each panel is shown the proportion of sessions in which decoding was significantly more accurate than when labels were shuffled. (c) Maximum decoding accuracy of trial features as a function of population size. Fitted lines correspond to the best fit second-order polynomial function, shown for visualization purposes. Dependence of decoding accuracy on population size was confirmed using Spearman’s rank correlation coefficient (previous choice: rho 0.5994, p-value 1.79e-4; previous outcome: rho 0.6466, p-value 3.61e-5; stimulus: rho 0.6558, p-value 2.6e-5; choice: rho 0.4481, p-value 7.9e-3; outcome: rho 0.7307, p-value 9.27e-7). Inset: Distribution of population sizes. The underlying data for this figure are available for download from 10.17632/5ms7gcb67j.1.