a7453a5f026fb6831d68bdc9cb0edcae-AuthorFeedback.pdf

Batch size has been an important component of past analyses. Regarding Comment 3,4 of Reviewer #2 and Comment 1 of Reviewer #5, we want to clarify we do not claim that our theory covers all the9 benefits of BN. The fast equilibrium conjecture only partially explains the benefits of BN. In Figure 7(b), the dotted red line (1/effective lr) increases by 10 times and then drops by12 approximately 10slowlyin40epochs. However, this is not observed in any of our settings, so it's not clear to us whether the heavy tail assumption holds for our setting.16