In this paper, some shortages of traditional BP learning algorithm are analyzed. To avoid these shortages, a modified BP learning algorithm is proposed. It is s hown that this algorithm is super-linearly convergent under certain conditions. This algorithm can overcome some shortages of traditional BP learning algorithm , and has the same order of computation complexity as the traditional BP algorit hm. Finally, two computing examples are given. Simulation results illustrate tha t this algorithm is highly effective and practicable.