We disagree on the judgement with our highest respect, due to the nontrivial technical differences and results

We thank all reviewers for their helpful and constructive comments. We'll further improve in the final version. In particular, our contributions are: (1) We introduce generalization bounds of learning algorithms on various losses, i.e. Besides, it's nontrivial to analyze the relationship between HL and RL, especially for the second inequality We'll add the discussions in the final version. We'll make the comparison and statements Below, we discuss the pros and cons of each one in detail.