PURPOSE: To provide a method for accelerating convergence of back propagation algorithm to be applied for providing learning process at multilayer perceptron type neural circuit network for recognizing the cluster of samples, for example.

CONSTITUTION: This method is provided with a 1st stage based on a basic conception for progressively increasing the learning ability of a circuit network by progressively adding new samples recognized by the multilayer perceptron type neural circuit network to the start cluster of learning samples, 2nd stage based on a conception for progressively increasing the learning ability of the circuit network by progressively adding samples not recognized before, and 3rd stage based on a conception for progressively increasing the learning ability of the circuit network through progressive collapsions in the meaning of assimilation between the recognized sample and the non-recognized sample and through the following presentation of them to the circuit network.