Sampling-based techniques can provide unbiased, high-confidence estimates of deep neural network operational accuracy at low cost, while also exposing many mispredictions to support model improvement.
