When Do Neural Networks Outperform Kernel
             Methods?

     Behrooz Ghorbani, Song Mei, Theodor Misiakiewicz, Andrea Montanari§

                         Abstract
      For a certain scaling of the initialization of stochastic gradient descent (SGD), wide
      neural networks (NN) have been shown to be well approximated by reproducing
      kernel Hilbert space (RKHS) methods. Recent empirical work showed that, for
      some classification tasks, RKHS methods can replace NNs wit ...