1 paper · 1 filter
David A. R. Robin, Kevin Scaman, Marc Lelarge
We present a framework to define a large class of neural networks for which, by construction, training by gradient flow provably reaches arbitrarily low loss when the number of par…