Tractable Structured Natural-Gradient Descent Using Local Parameterizations
            Wu Lin 1 Frank Nielsen 2 Mohammad Emtiyaz Khan 3 Mark Schmidt 1 4

             Abstract                  Finally, many robust or global optimization techniques em-
   Natural-gradient descent (NGD) on structured          ploy q(w) to smooth out local minima (Mobahi & Fisher III,
   parameter spaces (e.g., low-rank covariances)          2015; Leordeanu & Hebert, 2008; Hazan et al., 2016), where
   is computat ...