Difference FreeSel and ActiveFT

[zpge]

Excuse me. What is the difference between FreeSel and ActiveFT? It looks quite similiar.

[yichen928]

Hi, thanks for your interest.

ActiveFT is actually a follow-up work of FreeSel. We highlight several major differences of the two works:

1. Different settings. ActiveFT is targeted for the data selection for the supervised finetuning of pretrained models. FreeSel focuses on the from-scratch training of versatile models. However, our experiments also show that ActiveFT can also work in the from-scratch training setting.

2. Methods. ActiveFT adopts global representation format and aims to maximize the distribution similarity between the entire unlabeled pool and the selected subset. FreeSel adopts dense local representation and the goal is to cover diverse and representative local patterns in the selected subset.

3. Tasks. ActiveFT is mainly designed for classification task given its global representation format. FreeSel can work for different dense semantic tasks.

[zpge]

Hi, thanks for your interest.

ActiveFT is actually a follow-up work of FreeSel. We highlight several major differences of the two works:

1. Different settings. ActiveFT is targeted for the data selection for the supervised finetuning of pretrained models. FreeSel focuses on the from-scratch training of versatile models. However, our experiments also show that ActiveFT can also work in the from-scratch training setting.
2. Methods. ActiveFT adopts global representation format and aims to maximize the distribution similarity between the entire unlabeled pool and the selected subset. FreeSel adopts dense local representation and the goal is to cover diverse and representative local patterns in the selected subset.
3. Tasks. ActiveFT is mainly designed for classification task given its global representation format. FreeSel can work for different dense semantic tasks.

Thanks. Very clear.