Google DeepMind interview question

How can we optimize the focal loss over the Cross Entropy loss function ?