Training Deep Models Faster with Robust, Approximate Importance Sampling