티스토리 뷰

AI/Deep Learning

What's in a "Domain"

미남잉 2023. 3. 19. 21:42
728x90

 

What's in a "Domain"

Mathematically, joint distribution over inputs and outputs differs over domains 1 and 2

Pd1(X,Y)Pd2(X,Y)

 

예를 들어,

  • Content, whit is being discussed
  • Style, the way in which it is being discussed
  • Labeling Standards, the way thtat the same data is labeled

 

Types of Domian Shift

  • Covariate Shift: The input changes but not the labeling

Pd1(X)Pd2(X)

Pd1(Y|X)=Pd2(Y|X)

  • Concept Shift: The conditional distribution of labels changes (e.g. different labeling standards)

Pd1(Y|X)Pd2(Y|X)

 

Domian Adaptation

  • Train on many domians, or a high-resourced domain

  • Test on a low-resourced domain

  • Supervised or unsupervised adaptation

 

Domian Robustness

  • Train on many domains and do well on all of them

  • Robustness to minority domains
  • Zero-shot robustness to domains not in training data

 

출처: http://phontron.com/class/anlp2021/schedule/multitask.html

 

728x90

'AI > Deep Learning' 카테고리의 다른 글

Open Set Learning이란?  (0) 2023.04.14
Fine-Grained Image Classification이란?  (0) 2023.04.14
Zero-shot, Few-shot and Unsupervised Learning  (0) 2023.03.18
밑시딥3 정리  (0) 2023.03.08
Pretraining and Fine-tuning  (0) 2022.11.25