A fast fixed-point BYY harmony learning algorithm on Gaussian mixture with automated model selection

The Bayesian Ying–Yang (BYY) harmony learning theory has brought about a new mechanism that model selection on Gaussian mixture can be made automatically during parameter learning via maximization of a harmony function on finite mixture defined through a specific bidirectional architecture (BI-architecture) of the BYY learning system. In this paper, we propose a fast fixed-point learning algorithm for efficiently implementing maximization of the harmony function on Gaussian mixture with automated model selection. Several simulation experiments are performed to compare its effectiveness in automated model selection as well as its efficiency in parameter learning with other existing learning algorithms. The experimental results reveal that the performance of the proposed algorithm is superior to its counterparts in these aspects. Moreover, the proposed algorithm is further tested with three typical real-world data sets and successfully applied to unsupervised color image segmentation.