Fig. 4: Single-task neural manifolds are very dissimilar in PMd, moderately dissimilar in M1, and similar in S1.

a.\ The dimensionality of a dataset was estimated as the number of neural modes needed to explain more than 90% of the cumulative variance. The panel illustrates the process of comparing the dimensionality of single-task and all-task neural manifolds. Here, two 2D Tasks a and b generate a 3D all-task manifold. b.\ The difference between the dimensionality of all-task and single-task manifolds (Δ in dim) in PMd, M1 and S1 is large, moderate, and small, respectively. The bars show the histogram of single- (grayscale-coded) and all-task (yellow) manifold dimensionality across all sessions for Mk-Ek. The dots show the mean dimensionality of single- (gray) and all-task (yellow) manifolds (single-task: PMd: 3.46 ± 0.92; M1: 3.34 ± 0.48; S1: 2.20 ± 0.46; all-task: PMd: 8.71 ± 2.14; M1: 6.14 ± 0.69; S1: 3.42 ± 0.53). c The bars show the mean dimensionality of all-task manifolds across all sessions and both monkeys (PMd: 8.05 ± 0.44; M1: 5.33 ± 0.29 and S1: 4.78 ± 0.44; PMd vs. M1: p = 0.0005; PMd vs. S1: p = 0.0006; M1 vs. S1: p = 0.30; two-sided Wilcoxon signed rank test). Dots show values for each session and monkey (Mk-Ek: left column; Mk-Nt: right column). d. Difference (Δ) between dimensionality of all-task and single-task manifolds shown same as in c (PMd: 5.81 ± 0.38, M1: 2.66 ± 0.25 and S1: 2.11 ± 0.28; PMd vs. M1: p = 0.00024414; PMd vs. S1: p = 0.0001; M1 vs. S1: p = 0.15; two-sided Wilcoxon signed rank test). Values shown in (c, d) are means over 1000 repetitions of randomly selecting 32 channels from each cortical region to avoid bias due to different number of recording channels between regions and monkeys. e The panel illustrates the process of computing the alignment index. First, the neural activity of Tasks a and b are used to construct Task a and b single-task manifolds, respectively. Task a neural activity is then projected into the Task b manifold and, separately, into the Task a manifold. To obtain the alignment index, we divided the variance of the projection into Task b manifold by the variance of the projection into Task a manifold. f Bar plot shows the alignment index for both monkeys (dots on the left column, Mk-Ek; dots on the right column, Mk-Nt). Smaller alignment index of PMd compared to S1 shows that PMd single-task manifolds differ substantially, while S1 single-task manifolds are similar (n = 14; PMd: 0.38 ± 0.02; M1: 0.47 ± 0.02; S1: 0.56 ± 0.02; PMd vs. M1: p = 0.0005; PMd vs. S1: p = 0.0005; M1 vs. S1: p = 0.0005). Blue line shows the estimated noise contribution with the light blue tube showing the values not significantly different from noise contribution at p ≥ 0.05 (noise contribution: PMd: 0.891 ± 0.009; M1: 0.946 ± 0.005; S1:0.970 ± 0.005; measured value vs. noise contribution: PMd: p = 0.0005; M1: p = 0.0005; S1: p = 0.0005). Error bars: s.e.m.; *** p < 0.001; one-sided Monte Carlo permutation test. Source data are provided as a Source Data file.