Note: Below were experiments done at the early stage of this project. See THIS PAGE for more recent experiments where local shift leads to cleaner templates.

EM clustering from random initialization

The clustering is model-based. It does NOT require or rely on pairwise similarity meausre.