Reducing Overfitting in Deep Networks by Decorrelating Representations