Shunta Akiyama

This paper proves that a block coordinate descent algorithm can train deep neural networks to global minima under certain activation functions, extending to ReLU via architectural modifications.