It seems intuitive since ReLU is just a type of implicit regularization. Why would subsequent gradient descent help once you've achieved the benefit of throwing away the "outliers" or data beyond the threshold you want?