Statistical Physics of Deep Neural Networks: Initialization toward Optimal Channels