bigshanedogg / survey

2 stars 0 forks source link

[SimCLR v2] Big Self-Supervised Models are Strong Semi-Supervised Learners #2

Open bigshanedogg opened 2 years ago

bigshanedogg commented 2 years ago

Problem statement

  1. Unlabeled data를 간접적으로 활용하는 방법 외에 NLP에서 활용하는 것과 같이 더 직접적으로 활용하는 task-agnostic한 방법을 찾는다.
    • NLP에서의 활용: unsupervised pretrain -> supervised fine-tune
  2. unlabeled data를 활용해 general representation을 배우고 downstream task의 성능을 올리는 방법을 제안한다.

Baseline

Data details

Approach

Evaluation

Limitations

bigshanedogg commented 1 year ago

2006.10029.pdf