Recognition and geometrical on-line learning algorithm of probability distributions
Toshiaki Aida · 2000
An online learning algorithm for probability distributions is constructed in a reparameterization invariant form. It enables us to identify the distributions which transform from one to another by reparameterization. This is an essential property not only for pattern recognition problems but also for the property of 'information'. We can find the algorithm to be optimal, since conformal gauge reduces the problem to a noncovariant case.