Probabilistic Latent Variable Models as Nonnegative Factorizations
Figure 5
Example of the effect of the
entropic prior on a set of kernel and impulse distributions. If no constraint
is imposed, the information is evenly distributed among the two distributions
(left column), if sparsity is imposed on the impulse distribution, most
information lies in the kernel distribution (middle column), and vice verse if
we request a sparse kernel distribution (right column).