The research by Rohan Alur, Manish Raghavan, and Devavrat Shah introduces a novel framework for incorporating human expertise into algorithmic predictions. It emphasizes the potential of human judgment to improve the performance of predictive algorithms, particularly in areas where experts have access to subjective information that isn’t encoded in the algorithm’s training data. The researchers developed principled algorithms that selectively incorporate human feedback when it enhances any feasible predictor’s performance. They found that human judgment could greatly enhance algorithmic predictions on specific instances, with an example of an X-ray classification task where this approach improved nearly 30% of the patient population’s prediction.
Publication date: 1 Feb 2024
Project Page: https://arxiv.org/abs/2402.00793v1
Paper: https://arxiv.org/pdf/2402.00793
