“Using stale forward pass results for selection in Selective-Backprop can achieve an additional 26% acceleration in training by skipping forward passes of low-priority examples.”