Comparison of ST and SRTT trajectories in simultaneously recorded neurons.

A. Neural activity data projected on the PC1 (solid line, linearly detrended) and the correspondent sinusoidal fit (dotted line) during a trial of ST for the target interval of 650 ms. B. Similar to (A) for SRTT. Note that the strong periodic structure of the ST neural trajectory is lost during SRTT for the same population of cells. C. The MSE of the sinusoidal fits during ST (purple) is significantly smaller than during SRTT (green; 60 trials, two-sample t test = −6.78, p D. Radii of the neural trajectories during ST (purple, slope = 0.000087, constant = 0.055, R2 = 0.619, p R2 = 0.0172 and p = 0.489) as a function of target interval. E. Variability of the neural trajectories during ST (purple, data slope = 0.000037, constant = 0.028, R2 = 0.368, p R2 = 0.0005 and p = 0.903), and temporal variability of the monkeys’ produced intervals (gray, mean ± SD/2, data slope = 0.0009, constant = −0.003, R2 = 0.999, p F. Linear speed of neural trajectories during ST (purple, mean ± SD, slope = 0.0001, constant = 7.322, R2 = 0.0007, p = 0.896) and SRTT (green, mean ± SD, slope = 0.002, constant = 4.049, R2 = 0.354, p = 0.002) did not change as a function of target interval. G. Output of the time-delay neural network (TDNN, in blue) trained to decode the duration of produced intervals based on the PC1 neural trajectories (orange) during a target interval of 850 ms. Tapping times are shown in yellow. H. TDNN error, defined as the difference between the produced and the decoded interval, as a function of produced interval. TDNN predicted accurately the performance of the monkey on a trial-by-trial basis (the decoded mean was not statistically different from 0, t test = −0.5228, p = 0.6). Underlying data are available in https://doid.gin.g-node.org/d315b3db0cee15869b3d9ed164f88cfa/. a.u., arbitrary unit; MSE, mean square error; PC, principal component; SRTT, serial reaction time task; ST, synchronization task; TDNN, time-delay neural network.