Tight Bounds on the Smallest Eigenvalue of the Neural Tangent Kernel for
                Deep ReLU Networks

                 Quynh Nguyen 1 Marco Mondelli 2 Guido Montufar 1 3

             Abstract               We assume that the network has a single output, namely
                                nL = 1 and WL ∈ RnL1 . For consistency, let n0 = d.
    A recent line of work has analyzed the theoretical   Let gl : Rd → Rnl be the pre-activation feature map
    properties of deep neural network ...