Why did deep networks stop learning?

Vanishing and exploding gradients: why blame fades or blows up on its way back through a deep network, and the fixes that made deep learning work