Contradict the Machine: A Hybrid Approach to Identifying Unknown Unknowns

Machine predictions that are highly confident yet incorrect, i.e. unknown unknowns, are crucial errors to identify, especially in high-stakes settings like medicine or law. We describe a hybrid approach to identifying unknown unknowns that combines the previous algorithmic and crowdsourcing strategies. Our method uses a set of decision rules to approximate how the model makes high confidence predictions. We present the rules to crowd workers, and challenge them to generate instances that contradict the rules. To select the most promising rule to next present to workers, we use a multi-armed bandit algorithm. We evaluate our method by conducting a user study on Amazon Mechanical Turk. Experimental results on three datasets indicate that our approach discovers unknown unknowns more efficiently than state-of-the-art baselines.