Multi-Task Learning as Multi-Objective Optimization
本文将多任务学习作为多目标优化,总体目标是找到帕累托最优解。
现有的MGDA(multiple-gradient descent algorithm)缺陷:
(i) The underlying optimization problem does not scale gracefully to high-dimensional gradients, which arise naturally in deep networks.
(ii) The algorithm requires explicit co
原创
2022-01-25 23:23:55 ·
452 阅读 ·
0 评论