Fig 7. Separation of arm-specific subspaces relies upon unit-level segregation.

(A-C) Single session example of a PCA model trained to capture bi-hemispheric activity during left arm movements. Held-out testing data for 82 simultaneously recorded units were used. (A) Cumulative proportion of variance accounted for across the top 10 principal components. (B) For each component, the ratio of the explained variance between the two limbs. (C) For each component circled in red in 7B, the absolute values of the coefficient weights are plotted against the corresponding unit’s arm preference. Top row represents components 1–3; bottom row represents components 4–6. Positive arm preference values indicate right arm preferring units. (D) The component variance ratio for the two arms plotted against a coefficient-weighted average of the arm preferences for each unit in that component. Datapoints represent the top 5 principal components of left or right arm trained models across all sessions. Separate models for each phase are plotted in each column. Because these models include activity from both hemispheres, hands are referred to as “left” and “right” as opposed to “ipsi” and “contra”. Pearson correlation coefficient for each dataset is displayed in the red box. Top row monkey O, bottom row monkey W.