Defeating vanishing gradients in deep neural networks: Quotient Tree Arithmetic