Non-logarithmic information measures, α-weighted EM algorithms and speedup of learning
Yasuo Matsuyama · 2002
Starting from Renyi's /spl alpha/-divergence, a class of generalized EM algorithms called the /spl alpha/-EM algorithms of the WEM algorithms are derived. Merits of this generalization are found on speedup of learning, i.e., acceleration of convergence. Discussions include novel /spl alpha/-versions of logarithm, efficient scores, information matrices and the Cramer-Rao bound. The speedup is examined on Gaussian mixture learning systems.