Training (Overparametrized) Neural Networks in Near-Linear Time

The slow convergence rate and pathological curvature issues of first-order gradient methods for training deep neural networks, initiated an ongoing effort for developing faster mathitsecond-mathitorder optimization algorithms beyond SGD, without compromising the generalization error. Despite their remarkable convergence rate (mathitindependent of the training batch size n), second-order algorithms incur a daunting slowdown in the mathitcost mathitper mathititeration (inverting the Hessian matrix of the loss function), which renders them impractical. Very recently, this computational overhead was mitigated by the works of [ZMG19,CGH+19, yielding an O(mn2)-time second-order algorithm for training two-layer overparametrized neural networks of polynomial width m. We show how to speed up the algorithm of [CGH+19], achieving an tildeO(mn)-time backpropagation algorithm for training (mildly overparametrized) ReLU networks, which is near-linear in the dimension (mn) of the full gradient (Jacobian) matrix. The centerpiece of our algorithm is to reformulate the Gauss-Newton iteration as an ℓ2-regression problem, and then use a Fast-JL type dimension reduction to mathitprecondition the underlying Gram matrix in time independent of M, allowing to find a sufficiently good approximate solution via mathitfirst-mathitorder conjugate gradient. Our result provides a proof-of-concept that advanced machinery from randomized linear algebra -- which led to recent breakthroughs in mathitconvex mathitoptimization (ERM, LPs, Regression) -- can be carried over to the realm of deep learning as well.