Training MLP via the deterministic annealing EM algorithm
Yingjian Qi, Siwei Luo, Jian-Yu Li, Hong Tu · 2002
Supervised multi-layer perceptron (MLP) network is an important kind of artificial neural network model and have been used in many practical fields. In recent years, the EM (Expectation-Maximization) algorithm has been used to train the MLP network and has gotten good results. But the main problem associated with the algorithm is the local maxima problem. So we introduce the deterministic annealing method combined with the EM algorithm into MLP network to optimize the training method. In this paper we deduce the probability expression of the multi-output MLP model and give the DAEM training process. Experiment proves that our method is efficient.