Abstract: Combination of single output neural network （CSNN） has three merits: network structure being certain for a given classification task, understandable classification behavior, and easy parallelism, which overcomes the limitation of traditional BPNN, while classifying accuracy of CSNN is not as good as those of selected BPNNs In order to improve the generalization performance of CSNN, an error correcting output code is adopted as a distributed output representation, where each class is assigned a unique binary string of length n （codeword）, and these n binary functions are implemented by n SNN subnets A new case is classified by evaluating each of the n functions to generate an n bit output code, and assigned to the class whose codeword is closest, according to Hamming distance measure, to the generated string The behavior of this kind of CSNN can also be translated into a rule set Experimental results show that comparison with other supervised learning methods, CSNN with ECOC improves classifying accuracy greatly while retaining its original merits