Fixes That Fail: Self-Defeating Improvements in Machine-Learning Systems

Models in these systems are often developed and trained independently, which raises an obvious concern: Can improving a machine-learning model make the overall system worse?