Thanks for your great work in introducing GLMC. We had introduced Saddle-Long Tail work in NeurIPS 2022, which had demonstrated that using SAM with re-weighting based methods leads to improved generalization by escaping saddle points. As GLMC also uses the re-weighting loss, we integrated it with GLMC-2023. We find that SAM improves the performance of GLMC by around 1.5 - 2.0 % in various cases we tested, without any requirement of the second stage tuning. The details of the experiments are provided in the links below:
Please note that we only ran experiments once, hence there might be some variance in reported results. However, we are relatively confident that SAM will significantly improve the GLMC results. Hence, it would be great if you can include our contribution in GLMC. Thanks for the amazing work :).
Dear Authors,
Thanks for your great work in introducing GLMC. We had introduced Saddle-Long Tail work in NeurIPS 2022, which had demonstrated that using SAM with re-weighting based methods leads to improved generalization by escaping saddle points. As GLMC also uses the re-weighting loss, we integrated it with GLMC-2023. We find that SAM improves the performance of GLMC by around 1.5 - 2.0 % in various cases we tested, without any requirement of the second stage tuning. The details of the experiments are provided in the links below:
Code: https://github.com/val-iisc/Saddle-LongTail/tree/main/GLMC-2023
Results: https://github.com/val-iisc/Saddle-LongTail#results-with-glmc
Paper: arXiv
Please note that we only ran experiments once, hence there might be some variance in reported results. However, we are relatively confident that SAM will significantly improve the GLMC results. Hence, it would be great if you can include our contribution in GLMC. Thanks for the amazing work :).
Thanks