“The Selective-Backprop technique uses the output of a training example's forward pass to decide whether to compute gradients and update parameters for that example or to skip it.”