Bayesian Ying-Yang system and theory as a unified statistical learning approach. IV. Further advances

Abstract

A unified statistical learning approach called Bayesian Ying-Yang (BYY) system and theory has been developed by the present author in recent years. It functions as a general theory for supervised, unsupervised learning and semi-unsupervised extension for parameter learning, regularization, complexity selection, and architecture design. This paper reports several new advances: 1) the Ying dominated BYY learning are further discussed; 2) a general stochastic implementing procedure with detailed algorithms is proposed to overcome the difficulty encountered in the integral and summation operations such that the parameter learning and model selection become always implementable not only for the Yang dominated but also for the Ying dominated BYY learning; and 3) developments on BYY three-layer forward learning are provided, including new and simple criteria for the selection of the best hidden unit number and for the regularization of parameter learning.