The quadratic property of the L-MBFGS methods for training neural networks
Lin Zhao, Dali Wang, Yueting Yang · 2011
In this paper, we introduce the use of limited memory modified BFGS method (L-MBFGS) to improve the efficiency of training algorithms for feedforward neural networks. The quadratic termination property of L-MBFGS algorithm is given which is an important quasi-Newton property.