Non-negative matrix factorization(NMF) is formulated as a large-scale optimization problem with linear equality constraints.To get better performance

a hybrid combining simulated annealing(SA) and gradient-based algorithm is designed.The simulated annealing gradually produces better solutions with the gradient-based algorithm serving as an accelerator.Experimental results are presented to compare our method and the original method for learning NMF representation

which demonstrate the proposed method can learn basis images more spatially localized and perform better than the original one at later training stages.The comparison study on face reconstruction also shows that the proposed method leads to better results than PCA and the original NMF.