Gradient Surgery for Multi-Task Learning Tianhe Y u

We propose a form of gradient surgery that projects a task's gradient onto the normal plane of the gradient of