Abstract: Following the analysis of common model of neural networks and attribute theory and method, new attribute neural networks are proposed in this paper. The model saves data information to attribute nerve cell and joins function. Learning process has become simple and confirmative, and absolutely convergent. Graph property of attribute neural networks is discussed. A graph theoretic algorithm can be used to study sorter function of attribute neural networks. The equivalence of attribute neural networks and attribute coordinate space is proved, and a method of numerical value inference is provided for attribute inference.