search

median-research-group / LibMTL

A PyTorch Library for Multi-Task Learning
MIT License
1.94k stars 181 forks source link

Question about my understanding of aligned-MTL #70

Closed HelloWorldLTY closed 7 months ago

HelloWorldLTY commented 8 months ago

Hi, thanks for your great work. I have a quick question about the principle of this paper (aligned-MTL)

https://openaccess.thecvf.com/content/CVPR2023/papers/Senushkin_Independent_Component_Alignment_for_Multi-Task_Learning_CVPR_2023_paper.pdf

It seems that your implementation of algorithm 1 will return alpha, and here my understanding is we can directly replace the gradient matrix G with G\alpha. Is it different from your diagram?

image

If my understanding is incorrect, how to combine this component to my own dataset? It is a tarbular one and just simple regression tasks.

Baijiong-Lin commented 8 months ago

In our re-implementation, $\alpha$ will affect the update of task-specific parameters, following official implementation of Aligned-MTL.

HelloWorldLTY commented 8 months ago

OK, thanks a lot. I also have a question about multi-task learning. If my model performs well for 2 out of 3 tasks, but very bad in the third task (say, pearson correlation <0), and under this case, do you think using gradient alignment can significantly improve the model performance? Thanks.

Baijiong-Lin commented 8 months ago

I think you can try it, maybe it will be useful.

HelloWorldLTY commented 8 months ago

Thanks. Furthermore, could we directly use callback to set up early stop in the training process? Thanks.

Baijiong-Lin commented 7 months ago

No, LibMTL does not support callback.

Baijiong-Lin commented 7 months ago

Closed as no further updates.