Risk bound of priority ordered neural network with multi-weighted neurons

Neural networks are widely used in different fields. However, fixed architecture is difficult to use in practice for its pre-determined neurons and architectures. Constructive architecture is proposed in this paper for neural network based on idea of human's cognition where each neuron has its own priority number and evolution of learning process with time. Using prediction theory, risk bound of the architecture with multi-weighted neurons are analyzed. Experimental results show that the propose method outperforms the SVM method using limited samples. The proposed method is constructive and this work can provide a very useful method for neural network learning model.