Positive-Negative Momentum:
     Manipulating Stochastic Gradient Noise to Improve Generalization

               Zeke Xie 1 Li Yuan 2 Zhanxing Zhu 3 Masashi Sugiyama 4 1

             Abstract                 It is well-known that stochastic gradient noise (SGN) in
                                 stochastic optimization acts as implicit regularization for
   It is well-known that stochastic gradient noise       deep learning and is essentially important for both optimiza-
   (SGN) acts as impl ...