Is there a way to regularize the result by adjusting the loss function to prevent flipped normal? Or are we not doing so because it will be harder to differentiate?