Abstract: In pattern recognition, it is important to find a suitable way to decide a sample's class attribute according to its feature description. From the cybernetics point of view, the transformation is a black box. In this paper, we propose a new model integrated neural network with feedback, which can be implemented as a combine model for pattern recognition and is a closed-loop structure. We present the mathematical theorems for its absolute stability and the corresponding learning algorithm. Experiments on totally unconstrained handwritten numeral recognition have confirmed the proposed model's superiority.