After main results and the fundamentals of Bayesian Ying Yang (BYY) system and harmony learning is outlined, its ability on regularization and model selection is explained not only via its least complexity nature that minimizes the inner representation complexity, but also from an information transfer perspective. Moreover, with Ying-Yang pairs described in full distributions instead of simply in conditional distributions, this learning framework is further extended to a more general form for regularization and model selection in a more complicated BYY system, via controlling both the complexity of inner representation and the complexity of parameters of all the parts in the system.
展开▼
