Reviews: An Improved Analysis of Training Over-parameterized Deep Neural Networks

While this paper makes a nice contribution to an important problem, I am not sure if it is significant enough for the conference. The overall outline of the analysis follows closely that of [2], and the main new component is the improved gradient lower bound, which is largely based on previous ones in [2] and [16]. Although the improved analysis provides new insight and I find it useful, I do not feel that it will provide a big impact. The other technical contribution on improved trajectory length is also nice but again I feel that it is somewhat incremental. The results seem technically sound; the proofs all look reasonable although I did not verify them thoroughly.