From Two-Class Linear Discriminant Analysis to Interpretable Multilayer Perceptron Design