Ablation study of auxiliary losses？

[joeyz0z]

Hello， I was wondering about the role of auxiliary losses on each intermediate decoder layer. Can it help to accelerate the model convergence or for other purposes？ Thanks!

[ttengwang]

The intermediate loss splits the learning into multiple steps and may ease the learning process. I observed it improves both localization and captioning performance, but I didn't remember it helps convergence.

The design follows the DETR and Deformable-DETR and you may find more analysis in these papers.