Polylogarithmic width suffices for gradient descent to achieve arbitrarily small test error with shallow ReLU networks.

Ziwei Ji, Matus Telgarsky

Browse the full ICLR paper archive.