Hiroki Naganuma

全体概観

• MIT Press: これを読むべき！
• Environment and Distribution Shift
• Nature ref pdf
• Nature ref web
• Vector Inst のまとめ
• Vector Inst のまとめ pdf

• When Training and Test Sets are Different: Characterising Learning Transfer

• 自分でやったまとめ PDF

Shift の種類

• Covariate Shift (共変量シフト)
• Target Shift (Label Shift, Prior Probability Shift, Prevalance Shift / 流行シフト)
• Concept Shift

ref MIT Press

Nature ref pdf

Vector PDF

Covariate Shift

Label Shift (Prior Probability Shift / Target Shift)

• これは、prevalence shiftと同じ
• target shift ともいう？
• ‘Prior probability shift’ とも同じ？
  • ref
• Maximum Likelihood With Bias-Corrected Calibration is Hard-To-Beat at Label Shift Adaptation
  • slide

Shift への対応

Covariate Shift

• 重み付けする
• この重みはデータが、分布 p からきたか、 分布 q からきたかを学習する
• $\beta = p(y)/q(y) $　、　p は target distribution, q は source distribution

Label Shift

• 混合行列を計算する
• ref Precision

Other Reference

• Understanding Dataset Shift
• トップカンファレンスにおけるデータセットシフトと機械学習
• データセットシフトの学習理論

• 深層学習の不確実性 - Uncertainty in Deep Neural Networks -

• Covariate Shift – Unearthing hidden problems in Real World Data Science

Paper Reference

• A Critical Analysis of Distribution Shift : ICLR2021 Reject
  • まとめ

"Delving Deep into the Generalization of Vision Transformers under Distribution Shifts":

Paper: https://t.co/SC8rnrvFqY
Code: https://t.co/HAOwsK42aG

- We investigate the OOD generalization of vision transformers.
- We integrated domain adaptation techniques into transformers. pic.twitter.com/E5oZHPEo9q

— Ziwei Liu (@liuziwei7) June 19, 2021

彼らは以下の Shift の分類をしてる

• Pyramid Adversarial Training Improves ViT Performance

Dataset

See also OOD Datasets まとめ

Name	Shift Type	Reference
MNIST-C	Corruption Shift	Google Github
MNIST-R (RotatedMNIST)	Rotation? Shift	DomainBed
CMNIST (ColoredMNIST)	Correlation Shift	DomainBed
MNIST Background Random	Background Shift	LISA (Prev name of Mila) ‘s description Dataset
MNIST -> SVHN	Style Shift	Torchvision
Morpho-MNIST	Perturbation Shift	Github
ImageNet-Real	ID: Annotation Shift? (relabels the validation set of the original ImageNet in orderto correct labeling errors)	Github
ImageNet-V2	ID: Sample Selection Bias?	Github
ImageNet-C	Corruption Shift	Github
ImageNet-P	Perturbation Shift	Github
ImageNet-A	(places the ImageNet objects in unusual contexts or orientations)	Github
ImageNet-O	OOD (contains image concepts from outside ImageNet-1K)	Github
ImageNet-R	Style Shift (contains abstract or rendered versions of the object)	Github
ImageNet-Sketch	Style Shift	Github
Stylized-ImageNet	Texture Shift	Github