Can neural networks do better than the Vapnik-Chervonenkis bounds?