Zhipeng Zhou

We propose a new paradigm to explore the best trade-off through mode connectivity for multi-task learning.

We propose a new plug-and-play optimization method from the perspective of symmetry teleportation for mainstream multi-task learning approaches.