The proposed paper introduces a fast single-hidden layer feed-forward neural network with a discriminative information constraint (FSLFN-DI). Data is initially projected into a latent space based on the Fisher discriminant criterion. This projection matrix ensures that points of the same class cluster together and those of different classes scatter in the hidden latent space. A least square solution with the minimum norm is then used to learn the relationship between the hidden latent and the output for quick training. The experimental results on various industrial datasets demonstrate that FSLFN-DI outperforms traditional neural networks in terms of faster learning and competitive performance.