I just read the Inception-v3 paper, and its training code released by the authors. And I found that when do RMSProp optimization, the authors used epsilon=1. However, to my knowledge, people usually used 1e-10 or some small values and Tensorflow set the default value to 1e-10.
So my question is how will the epsilon influence the optimization? Why we need a epsilon=1 here? Thank you!