A technique using dropout during inference to estimate model uncertainty by sampling multiple predictions.