<FONT face=Verdana>This paper puts forward to the concept of loosely and tightly grouping-cascaded BP network model to solve the training problem of high dimension BP neural network with limited small samples. After defining the equivalence definition and relative theorem of BP neural network

the grouping-cascaded model which is proved to be equivalent to BP network is constructed and the required number of training samples between them is compared. Then this method is applied in the evaluation of the network security. Finally

the feasibility and validity of the proposed grouping-cascaded BP network model is verified with simulation results.