What if Neural Networks had SVDs?

     Alexander Mathiasen      Frederik Hvilshj             Jakob Rdsgaard Jrgensen
             Anshul Nasery  
                                       Davide Mottin

                         Abstract
       Various Neural Networks employ time-consuming matrix operations like matrix
       inversion. Many such matrix operations are faster to compute given the Singular
       Value Decomposition (SVD). Techniques from [10, 17] allow using the SVD in
     ...