티스토리 뷰

AI/주워들은 것들

CI-UDA 읽은 후기

아인샴 2024. 3. 30. 20:57

(1)

2024.03.25 - [AI] - 0325 papers + CI-UDA

(2)

2024.03.28 - [AI] - Class-Incremental Unsupervised Domain Adaptation via Pseudo-Label Distillation(2024)

 

공식까지 빠지지 않고 처음부터 끝까지 다읽은 논문은 처음이다. 그런 의미에서 논문에 있는 말을 번역한 것이 아닌 내가 이해한 바를 따로 정리해 보려고 한다. 

 

 

class-level 에서는 confident  한 target sample과 src domain 간의 feature representation 을 맞추고(align), instance-level에서는 unconfident target sample의 robust featrue representation을 활용하는 Contrastive Algignment를 제안한다. 이를 통해 domain adaptation의 performance을 향상시키고자 한다.
  • 여기서 class-level과 instance-level이 헷갈렸는데 그림으로 치면 이런 것 같다. 

여전히 이게 맞는지 확신이 안가는 부분도 좀 있음(previous step이라 함은 class가 늘어나서 그런 것을 말하는게 아닌가 싶다.)

 

그런데 시간이 부족해 일단 중요단어 10개만 뽑아보자. 

 

ProCA - memory bank
biased source information
negative trnsfer 
Continual Domain Adaptation
catatrophic forgetting

contrastive learning - contrastive alignment - adversarial learning(feature extractor VS discriminator) 
Prototypes(class-level prototype)
domain-invarint knowlege
pseudo Label distillation
feature centroid