Joint Training for Selective Prediction

To problems is how best to provide automated support avoid confusion with the main classifier in the SP for decisions that humans find laborious, have difficulty setting (CL), we will refer to a deferral classifier making in a consistent and therefore fair manner, as a deferral policy (DP). Often, a deferral policy and that have high social impact. In areas such is trained on a dataset distinct from the one used as education, human expertise can be best utilized to train the CL. The key issue is that the supervision when experts are relieved of more routine decisions signal for the deferral policy depends on the that can be reliabily automated. Yet, concerns have CL producing its decisions. As a result, all previous been raised about how machines and humans can methods have relied on a previously trained become better collaborators in education (Cardona CL. Here we jointly train the classifier and the deferral et al., 2023). Educators can learn more about students' policy, which we refer to as Joint Training misconceptions when they answer questions for Selective Prediction (JTSP). We find that JTSP in their own words, so automated support could outperforms other methods, and also improves the shift human effort from grading towards design of accuracy of each of its modules (CL and DP).