Abstract: A feedback procedure neural network model(FPNN) is proposed. The FPNN has three layers, and its hidden layer and output layer are composed of procedure neurons. The input layer accomplishes continuous signal input, while the hidden layer accomplishes input signal aggregation in space and transfers the input signals to the output layer. Then the hidden layer transfers its own output to the input layer, both point by point. The output layer accomplishes output signal aggregation in both space and time, and fulfills system output. A learning algorithm is pre sented based on Walsh conversion of weight function. Simulation experiment proves the availability and effectiveness of the model and algorithm.