Figure 7

Stability of NBGNet’s predictions for multiple days. (a,b) Histogram of RMSE (left) at Day 1 (blue) and 16 (purple) for the forward (a) and the inverse (b) model. Scatter plot of average RMSE (right) showing no significant difference (error bars, s.e.m.; n = 16 and 157 for (a) and (b); p = 0.29 and 0.07 for a and b using one-way ANOVA test). (c,d) Beta correlation (left) at Day 1 (blue) and 16 (purple) for the forward (c) and the inverse (d) model. Scatter plot of average beta correlation, where black solid line is obtained by averaging over the channels and black dashed line left the poorest channel out (error bars, s.e.m.; n = 7 and 16 for (c) and (d)). p = 0.32 and < 0.05 for c and d using one-way ANOVA test. While there are some significant decreases for the inverse model, the effect size is small. (e,f) Scatter plot of PSI versus PLI (left) at Day 1 (blue circle) and 16 (purple triangle) for the forward (e) and the inverse (f) model. Stacked bars (right) demonstrate the percentage of predictions locating in each section. (g,h) Temporal correlation averaging across days (upper), where colored segments represents stronger correlation as compared with the grey counterparts. Bar plot of average correlation (lower) exhibiting stable performance (error bars; s.e.m.; n = 68, 49, 128, 78, 135 at Day 1, 2, 4, 12, 16). *p < 0.05, **p < 0.01, ***p < 0.001 using two-sided Wilcoxon’s rank-sum test. n.s. indicates no significant difference. (i) Bar plots showing the classification accuracy of linear classifier to predict the target direction (error bars, s.d.; n = 5). Solid line represents the average performance across days. (j) Performance comparisons between the NBGNet, GRU-RNN, and the sphere head model. **p < 0.01, ***p < 0.001 using two-sided Wilcoxon’s rank-sum test.