MDA projection subspace for the cortical arm movement simulated data set.

<p>A) First dimension yields separation between the black cluster (corresponding to the no-movement states) and the responses to arm movement at different angles (green for the 45°, blue for the 90°, magenta for the 135° and cyan for the 270° classes). Second dimension separates the forward movement clusters (45°, 90° and 135°) from backward movement cluster (270°). B) Third dimension separates the left and right directions (green 45° from magenta 135°), while separation between forward (blue 90°) and lateral-forward (green 45° from magenta 135°) is attempted in the last dimension. The generalization performances for these dimensions correlate with the magnitude of the corresponding eigenvalues sequence generated by MDA: 0.74, 0.16, 0.06 and 0.03 (sum of all eigenvalues is normalized to 1). C) Smoothed weight distributions on the first dimension indicate that neurons responsive to all movements have the largest contribution to this dimension (thick blue line) followed by broadly tuned neurons (thick cyan line) and sharply tuned neuron (green line). Neurons sharply tuned to a specific angle (thin red line) and non-responsive (thin black line) units have very small contributions. Stimulating angles (green 45°, blue 90°, magenta 135° and cyan 270°) are listed as colored stars symbols on the top of the plot. D) In the second dimension the broadly tuned neurons have the largest contributions, followed by sharply tuned and angle-specific neurons, with the rest of the units having negligible contributions. E) Third dimension, which separates left from right, relies more on the sharply tuned units and less on the broadly tuned units, with minimal contributions from the rest of the neurons. F) Inspection of the fourth dimension curves indicate that the very sharply tuned neurons are used in the attempt to separate forward from oblique movement.