Abstract | ||
---|---|---|
A class of non-logarithmic likelihood ratio is considered and is applied to learning of neural networks including hierarchical experts. Such a likelihood ratio is based on an ct-logarithm which contains the usual logarithm as a special case. This generalized logarithm is defined through a discussion of the a-divergence which includes the Kullback-Leibler number as a special case. It is found that the usage of such a generalized logarithm on the likelihood ratio is equivalent to a prior probability weight. Then, this prior weighting is derived for learning on:neural networks of expert mixtures. Both of gradient ascent maximization and EM learning are discussed. The prior weighting is understood as speed-up and stabilization on the learning. |
Year | Venue | Field |
---|---|---|
1997 | PROGRESS IN CONNECTIONIST-BASED INFORMATION SYSTEMS, VOLS 1 AND 2 | Neural network learning,Computer science,Artificial intelligence,Prior probability,Machine learning |
DocType | Citations | PageRank |
Conference | 0 | 0.34 |
References | Authors | |
0 | 3 |
Name | Order | Citations | PageRank |
---|---|---|---|
Yasuo Matsuyama | 1 | 60 | 16.41 |
S. Furukawa | 2 | 0 | 0.34 |
T. Ikeda | 3 | 0 | 0.34 |