Revealing the Structure of Deep Neural Networks via Convex Duality

                       Tolga Ergen 1 Mert Pilanci 1

            Abstract                where two-layer ReLU networks with the minimum Eu-
                                clidean norm solution and zero training error are proven
   We study regularized deep neural networks         to fit a linear spline model in 1D regression. In addi-
   (DNNs) and introduce a convex analytic frame-       tion, a recent series of work (Pilanci  ...