Why Multi-Task Learning Reduces Accuracy: The Mathematical Inevitability of Simultaneous Optimization for Multiple Objectives

If foundation models like CNNs and BERT are strong at multiple tasks, then training them on multiple tasks simultaneously ...