NeurIPS 2020

Probably Approximately Correct Constrained Learning

Meta Review

The authors propose an extension to PAC learning by considering explicitly constraints that are imposed in particular settings. These could be, for example, fairness constraints or robustness constraints. The theory and the experimental set up are technically correct. The work is of high relevance to the machine learning community as it incorporates into a nice formalism the constraints that are often imposed on the learning algorithms.