Excited to share PCGrad, a super simple & effective method for multi-task learning & multi-task RL: project conflicting gradients
On Meta-World MT50, PCGrad can solve *2x* more tasks than prior methods
arxiv.org/abs/2001.06782
w/ Tianhe Yu, S Kumar, Gupta, ,
Conversation
With this and Orthogonal Weight modification, nice to see some simple linear algebra ideas scaling up to deep RL
1
3
That reminds me of this
Quote Tweet
Here is a blog post about auxiliary tasks in machine learning. I test a simple approach to mitigate negative transfer and the first experimental results are encouraging (@TensorFlow notebooks provided and directly usable in @GoogleColab)
vivien000.github.io/blog/journal/l
2
Cool work. Would explore if conflicting gradients can be used for new direction in the task space
