Skip to main content
. Author manuscript; available in PMC: 2015 Aug 1.
Published in final edited form as: Neuroimage. 2014 Mar 28;96:245–260. doi: 10.1016/j.neuroimage.2014.03.048

Figure C.13.

Figure C.13

Left) The gradient, g→, has two components: one for contrastive divergence, g→CD, the other from L1 regularization, g→L1 . If the dot product of g→Tg→CD>g→Tg→L1, then contrastive divergence dominates. We measure the dot products at iterations W away from the solution, W* . Right) Dot products of the contrastive divergence and L1 gradients, g→CD and g→L1 , onto the total gradient, g→, for real fMRI (top) and simulation (bottom) data across step distance from solution with SMs thresholded at 1 standard deviation. The L1 gradient and CD gradients are equal near the solution, but CD quickly dominates farther away. This indicates that while L1 as a stronger role close to the solution, the overall gradient is dominated by CD.