Fig. 4: Linear classification of population activity predicts animal’s choice and RT.

a Schematic to illustrate the linear decoding method using population activity. Clouds of hypothetical population responses corresponding to different conditions (left or right choices) are linearly discriminated in high dimensional neural activity space. b Classification accuracy of decoders trained on 23 simultaneously recorded neurons from one example session. Choice decoding corresponds to classification between left versus right choices; sensory decoding corresponds to classification between high versus low click rate trials; gray lines show performance on data with shuffled trial labels. Mean ± s.e.m. across 100 cross-validated samples. c Decoding accuracy for all sessions (n = 9), similar to b. Thin lines, individual sessions; thick lines, mean across sessions. d Schematic to illustrate leave-one-out cross-validation. e Left, delay durations binned over each 20 percentile of response times (RT) across sessions (n = 9). Gray dots, individual sessions; black dots, mean across sessions. Right, delay activity choice decoding accuracy binned over each 20 percentile of RT. Light pink dots, individual sessions; Dark pink dots, mean across sessions. r, Pearson’s correlation coefficient, P values report the significance of t tests.