Globally optimal learning rates for multilayer neural networks