This paper investigates when does curriculum learning~(CL) works in the background of computer vision tasks.
The motivation is CL has been used and shown to be effective for language modeling and neural machine translation, while whether CL is good for other machine learning tasks is unknown.
The findings of this paper include:
CL makes no difference with CV datasets when models get trained fully (e.g. 100 epochs)
CL does make difference when trained in limited epochs or dataset with noise (which is as same as the LM scenery)
This paper investigates when does curriculum learning~(CL) works in the background of computer vision tasks. The motivation is CL has been used and shown to be effective for language modeling and neural machine translation, while whether CL is good for other machine learning tasks is unknown.
The findings of this paper include: