Title :
Non-logarithmic information measures, α-weighted EM algorithms and speedup of learning
Author :
Matsuyama, Yasuo
Author_Institution :
Dept. of Electr., Electron. & Comput. Eng., Waseda Univ., Tokyo, Japan
Abstract :
Starting from Renyi´s α-divergence, a class of generalized EM algorithms called the α-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 α-versions of logarithm, efficient scores, information matrices and the Cramer-Rao bound. The speedup is examined on Gaussian mixture learning systems
Keywords :
convergence of numerical methods; information theory; optimisation; α-weighted EM algorithms; Cramer-Rao bound; Gaussian mixture learning systems; Renyi´s α-divergence; WEM algorithms; convergence acceleration; generalized EM algorithms; information matrices; learning speedup; nonlogarithmic information measures; Acceleration; Convergence; Electric variables measurement; Electrochemical machining; Entropy; Information theory; Intersymbol interference; Learning systems; Linear matrix inequalities; Velocity measurement;
Conference_Titel :
Information Theory, 1998. Proceedings. 1998 IEEE International Symposium on
Conference_Location :
Cambridge, MA
Print_ISBN :
0-7803-5000-6
DOI :
10.1109/ISIT.1998.708990
