A K-Means implementation for KL-Divergence (instead of squared euclidean distance)
See:
- http://jmlr.csail.mit.edu/papers/volume6/banerjee05b/banerjee05b.pdf
- http://mark.reid.name/blog/meet-the-bregman-divergences.html
- http://en.wikipedia.org/wiki/Bregman_divergence
To understand why this works. I'll add a better description later.