Training Binary Perceptrons by Clipping