A case for soft loss functions