Figure 3.

Nonlinear decoders leverage laterality information to disentangle effectors. (a) Offline single-bin decoding on unimanual data. Neural activity was binned (20-ms bins) and truncated to 400 ms movement windows (300–700 ms after go cue). Linear ridge regression (RR) and a densely connected feed forward neural network (FFN; single layer, 512 units) were trained, using five-fold cross-validation, to decode left and right cursor velocities. Sample 8 s held-out snippets of decoded x-direction velocity traces are shown. (b) Each bar indicates the offline decoding performance (Pearson correlation coefficient) for the RR and FFN decoders across the x- and y-direction velocity dimensions, separated by left hand (purple bars) and right hand (blue bars). Striped bars indicate data where the laterality dimension was removed. The FFN outperformed the LD in decoding movements across all dimensions. Removal of the laterality dimension did not affect LD performance but did reduce FFN performance. (c) Cursor jitter is quantified as the ratio of average cursor speed during rest periods to that during movement periods. A rest period is defined as the period in which the other cursor should be active. Lower ratios indicate less cursor jitter (or more cursor stillness) while the other cursor is active. The FFN outperformed the LD, maintaining a more stable left (non-dominant) cursor position in comparison. The laterality dimension was useful to the FFN in reducing cursor jitter; again, laterality did not affect the LD. (d) Simulated neural activity during unimanual movement was generated for different directional tuning correlation values between the hands and different laterality dimension sizes. Each (i,j) cell of a matrix indicates the decoding performance (Pearson correlation coefficient) for a synthetic dataset with correlation i between hands and a laterality dimension size of j. (e) Cursor jitter for the simulated data in panel d is shown. The FFN leveraged the laterality dimension for improved decoding performance and less cursor jitter as tuning between the hands became more correlated. The LD was unable to use the laterality information to distinguish between the hands.