High-dimensional data are often too complex to be classified. K-L transformation is an effective dimension reduction method. However its result is not satisfactory in large-set pattern classification. In this paper a novel nonlinear dimension reduction method is presented and analyzed. The transform is achieved through a multi-layer feed-forward neural network trained with K-L transformation result. Experimental results show that this method is more effective than K-L transformation being applied in large-set pattern classification such as Chinese character recognition.