The purpose of this study is the suggestion of a self-organized back-propagation algorithm for robust principal component analysis. The self-organizing algorithm that discriminates the influence of data automatically is applied to learning of a sandglass type neural network.