For　the　problem　that　the　fully　coupled　BP　neural　network　suffers　the　slow　convergence　rate　to　solve　the　large　scale　complex　problems,　a　structure　model　of　function-dividing　BP　neural　network　architecture　is　presented.　By　using　the　physical　characteristics　of　the　RBF　neurons,　the　input　sample　space　is　decomposed,　and　different　sub-samples　space　is　sent　to　different　sub-module　of　BP　neural　network　to　learn　automatically.　Compared　with　the　fully　coupled　BP　neural　network,　the　searching　space　of　weight　in　the　learning　process　of　neural　network　is　reduced,　the　learning　speed　and　network＇s　generalization　performance　are　improved,　and　the　characteristics　of　the　human　brain　in　the　learning　proces　of　knowledge　accumulation　are　reflected.　Experiments　of　3D　Mexican　hat　function　approximation　and　two-spiral　classification　show　that　the　neural　network　of　function-dividing　BP　neural　network　can　solve　the　problem　that　the　fully　coupled　BP　neural　network　can　not　solve　perfectly.