On the Expressive Power of Deep Learning: A Tensor Analysis