Fig. 2: Full kinematics encoding model predicts single-unit activity.
a–h Left: monkey TY. Right: monkey MG. a Mean ± sem AUC for the full kinematics model, all lead-lag sets (n = 500 per set). The trajectory models range from 300 ms to 500 ms in duration, and from lead-heavy samples on the left to lag-heavy samples on the right. The arrow indicates the best-performing lead-lag set. Solid lines above the data denote lead-lag sets for which AUC distributions were not significantly different than the best model (p > 0.05, one-sided sign test with Bonferroni correction). b Scatterplot of the trajectory and velocity model AUC values for each unit, averaged over 500 train/test splits. Each unit’s hue corresponds to average in-weight in the FN, which will be explained in further detail in Fig. 3. Units above the unity line were predicted better by the trajectory model. P-values are the result of a one-sided sign test. The trajectory model predicted activity better than the velocity model (TY: p ≈ 0.0; MG: p = 5.9 × 10−14; one-sided sign test). c The full kinematics model predicted activity better than the short kinematics model (TY: p = 2.1 × 10−24; MG: p = 7.7 × 10−4; one-sided sign test). In (b) and (c), the asterisk indicates that we filtered out units that were not significantly tuned to the trajectory, leaving 163/175 units for TY and 59/73 units for MG. d The full kinematics model outperformed the trajectory model due to the inclusion of average position terms (TY: p = 1.3 × 10−37; MG: p = 2.9 × 10−19; one-sided sign test). e Summary of model performance for 163 units (TY) and 59 units (MG). Center line indicates the median value, whiskers incorporate the middle 95% of data, and circles show the 2.5% of data at each end of the distribution. The full kinematics model produced AUC values significantly higher than all other models (TY: in addition to p-values presented in (c, d), p ≈ 0.0 for full kinematics vs. velocity, p ≈ 0.0 for full kinematics vs. total shuffle and p ≈ 0.0 for full kinematics vs trajectory shuffle; MG: p = 2.8 × 10−19, p = 1.8 × 10−11, and p ≈ 0; one-sided sign test). f The preferred trajectory pathlets for four units with high AUCs. Each of 500 train/test splits is shown in blue and red corresponding to lead and lag movements, and the average pathlet is shown in black. g The pathlets for four units with the lowest AUCs. h Histograms of the Pearson correlation between pathlets for all pairs of units. Source data are provided as a Source Data file.
