Abstract: A new neural network learning algorithm based on Bayesian method （BEM algorithm） is presented. One disadvantage of the standard EM algorithm is its convergence to local minimum. The BEM algorithm overcomes the disadvantage by analysing the reasons. Furthermore, BEM avoids the EM’s overfitting problem as well as its disability of generalization due to only responding to single patterm. Experimental result shows the BEM’s correctness and validity. The BEM has academic values of contributing to the research and development of EM algorithm.