paper
cdancette opened this issue · comments
Hi, this is a fantastic list of references!
Is there any chance you will consider adding our work too (https://www.aclweb.org/anthology/2020.acl-main.727.pdf)? Our study basically shows that the visual grounding methods (HINT/SCR) improve accuracy on VQA-CP through regularization effects as opposed to improving visual grounding. Example: SCR with irrelevant cues can achieve 49.2% accuracy (comparable to the accuracy achieved with relevant cues).
We also present a regularizer designed to simply degrade the training accuracy. It happens to achieve 48.9% accuracy, providing further evidence as to how the improvements stem from the regularization effects.
@erobic Thank you for your kind words !
I forgot to add your paper, but I'll add this soon ! It looks very interesting. Btw I am familiar with your other papers on VQA, and I find them very interesting ! Thanks for your work.
Oh, that's so great to hear! I am familiar with many of your works too, especially RUBi and the wonderful bootstrap framework.
Thanks a lot for considering the ACL paper! :-)
@erobic I just added it under the name ESR (embarassingly Simple Regularizer)
Great, thank you for including it!
ESR is really just to showcase how OOD benchmarks can be hacked without any core improvement.
Here are the rest of the fields for it: Yes/No: 69.8, Num: 11.3, Others: 47.8, No valset.
Thanks again!