Title :
Dual form back propagation on the EM algorithm
Author :
Hu, Hong ; Shi, Zhongzhi
Author_Institution :
Key Lab. of Intell. Inf. Process., Inst. of Comput. Technol., Beijing, China
Abstract :
Vladimir N. Vapnik. (1998) pointed out that maxlikelihood functions in EM algorithms are just a special risk function. Based on this opinion, a novel EM algorithm uses a risk function differ with maxlikelihood functions, in stead, a risk formula based on the least square method is used. The gradient descending approach should be used in such kind approaches. Such kind EM algorithms can estimate the parameters of a random model from both labeled and unlabeled samples, and are suitable for semi-supervised learning.
Keywords :
backpropagation; expectation-maximisation algorithm; gradient methods; learning (artificial intelligence); least squares approximations; parameter estimation; EM algorithm; dual form back propagation; gradient descending approach; least square method; maxlikelihood function; parameter estimation; risk formula; semisupervised learning; special risk function; Error analysis; Gaussian distribution; Information processing; Laboratories; Least squares methods; Maximum likelihood estimation; EM; back propagation learning; mixture Gaussian; semi-supervised learning;
Conference_Titel :
Granular Computing (GrC), 2011 IEEE International Conference on
Conference_Location :
Kaohsiung
Print_ISBN :
978-1-4577-0372-0
DOI :
10.1109/GRC.2011.6122604
