No bad local minima: Data independent training error guarantees for multilayer neural networks | Read Paper on Bytez