[DL輪読会]"CyCADA: Cycle-Consistent Adversarial Domain Adaptation"&"Learning Semantic Representations for Unsupervised Domain Adaptation"
7 likes2,631 views
1. Two papers on unsupervised domain adaptation were presented at ICML2018: "Learning Semantic Representations for Unsupervised Domain Adaptation" and "CyCADA: Cycle-Consistent Adversarial Domain Adaptation". 2. The CyCADA paper uses cycle-consistent adversarial domain adaptation with cycle GAN to translate images at the pixel level while also aligning representations at the semantic level. 3. The semantic representation paper uses semantic alignment and introduces techniques like adding noise to improve over previous semantic alignment methods.
![1
DEEP LEARNING JP
[DL Papers]
http://deeplearning.jp/
"CyCADA: Cycle-Consistent Adversarial Domain Adaptation"
& "Learning Semantic Representations for Unsupervised
Domain Adaptation" (ICML2018 )
Presentater: Kei Akuzawa, Matsuo Lab. M2](https://image.slidesharecdn.com/20180803dlakuzawa-180803003349/85/DL-CyCADA-Cycle-Consistent-Adversarial-Domain-Adaptation-Learning-Semantic-Representations-for-Unsupervised-Domain-Adaptation-1-320.jpg)
![•
•
•
• ICML2018
• [Hoffman+] CyCADA: Cycle-Consistent Adversarial Domain Adaptation
• [Xie+] Learning Semantic Representations for Unsupervised Domain Adaptation](https://image.slidesharecdn.com/20180803dlakuzawa-180803003349/85/DL-CyCADA-Cycle-Consistent-Adversarial-Domain-Adaptation-Learning-Semantic-Representations-for-Unsupervised-Domain-Adaptation-2-320.jpg)
![:
• Source
q(y|x)
( !"[$% & ' ] ≠ !*[$* & ' ] )
• z( )
p(z|x) p(y|z)
• z ?:
• MMD,
x
z
dy
z](https://image.slidesharecdn.com/20180803dlakuzawa-180803003349/85/DL-CyCADA-Cycle-Consistent-Adversarial-Domain-Adaptation-Learning-Semantic-Representations-for-Unsupervised-Domain-Adaptation-7-320.jpg)
![1.
• Title:
Learning Semantic Representations for Unsupervised Domain Adaptation
• Authors:
Shaoan Xie, Zibin Zheng, Liang Chen, Chuan Chen
• Info:
ICML2018 accepted (oral)
• Abstract:
[ , ]
Noisy
Centroid Noise
[ , ]](https://image.slidesharecdn.com/20180803dlakuzawa-180803003349/85/DL-CyCADA-Cycle-Consistent-Adversarial-Domain-Adaptation-Learning-Semantic-Representations-for-Unsupervised-Domain-Adaptation-12-320.jpg)
![[ , ]
Saito+ 2018](https://image.slidesharecdn.com/20180803dlakuzawa-180803003349/85/DL-CyCADA-Cycle-Consistent-Adversarial-Domain-Adaptation-Learning-Semantic-Representations-for-Unsupervised-Domain-Adaptation-13-320.jpg)
![• Semantic Alignment [saito+ 2017]
1.
2.
3. (w/ )
4.
• (Noisy)
•
•
• Centroid](https://image.slidesharecdn.com/20180803dlakuzawa-180803003349/85/DL-CyCADA-Cycle-Consistent-Adversarial-Domain-Adaptation-Learning-Semantic-Representations-for-Unsupervised-Domain-Adaptation-16-320.jpg)
![• [ ]
(Semantic Alignment )
• Semantic Alignment Noisy
• Noise](https://image.slidesharecdn.com/20180803dlakuzawa-180803003349/85/DL-CyCADA-Cycle-Consistent-Adversarial-Domain-Adaptation-Learning-Semantic-Representations-for-Unsupervised-Domain-Adaptation-20-320.jpg)
![•
• Semantic Alignment
•
• c.f. ( ) [Jo and
Bengio, 2017]
• ( )
•
•
• Cycle-GAN
•
• Cycle-GAN
•](https://image.slidesharecdn.com/20180803dlakuzawa-180803003349/85/DL-CyCADA-Cycle-Consistent-Adversarial-Domain-Adaptation-Learning-Semantic-Representations-for-Unsupervised-Domain-Adaptation-22-320.jpg)
![•
•
•
• [Saito+ 2018b]:
• [Hosseini-Asl+ 2018]:](https://image.slidesharecdn.com/20180803dlakuzawa-180803003349/85/DL-CyCADA-Cycle-Consistent-Adversarial-Domain-Adaptation-Learning-Semantic-Representations-for-Unsupervised-Domain-Adaptation-32-320.jpg)
[DL輪読会]"CyCADA: Cycle-Consistent Adversarial Domain Adaptation"&"Learning Semantic Representations for Unsupervised Domain Adaptation"
- 1. 1 DEEP LEARNING JP [DL Papers] http://deeplearning.jp/ "CyCADA: Cycle-Consistent Adversarial Domain Adaptation" & "Learning Semantic Representations for Unsupervised Domain Adaptation" (ICML2018 ) Presentater: Kei Akuzawa, Matsuo Lab. M2
- 2. • • • • ICML2018 • [Hoffman+] CyCADA: Cycle-Consistent Adversarial Domain Adaptation • [Xie+] Learning Semantic Representations for Unsupervised Domain Adaptation
- 3. • : Source Target Source Target • : • Source: (X_s, Y_s) • Target: • Unsupervised Domain Adaptation: (X_t) <- • Supervised Domain Adaptation: (X_t, Y_t)
- 5. • : • • • • Bengio ……(Talk at the ICML'2018 Workshop on Learning with Limited Labels, July 13th, 2018.) • Current ML theory is strongly dependent on the iid assumption • Real-life applications often require generalizations in regimes not seen during training • Humans can project themselves in situations they have never been (e.g. imagine being on another planet, or going through exceptional events like in many movies) ,
- 6. • • Ganin+ 2016 • Tzeng+ 2017 • Saito+ 2018 • • Taigman+ 2017 • Shrivastava+ 2017 • Hosseini-Asl+ 2018 Ganin+ 2016 Shrivastava+ 2017
- 7. : • Source q(y|x) ( !"[$% & ' ] ≠ !*[$* & ' ] ) • z( ) p(z|x) p(y|z) • z ?: • MMD, x z dy z
- 8. : Ganin+ 2016 • : • Discriminator (z) • Encoder Discriminator (z) • Min-Max (z)
- 9. !! • • Fair Prediction: (Domain) • Style Transfer: • Domain Generalization:
- 10. : • Target y • • • Unsupervised Image Translation (X_s) (X_t^') (X_t^', y_s)
- 11. • Learning Semantic Representations for Unsupervised Domain Adaptation • + ( ) • CyCADA: Cycle-Consistent Adversarial Domain Adaptation • + ( )
- 12. 1. • Title: Learning Semantic Representations for Unsupervised Domain Adaptation • Authors: Shaoan Xie, Zibin Zheng, Liang Chen, Chuan Chen • Info: ICML2018 accepted (oral) • Abstract: [ , ] Noisy Centroid Noise [ , ]
- 13. [ , ] Saito+ 2018
- 15. : C !! ( C) Semantic Alignment
- 16. • Semantic Alignment [saito+ 2017] 1. 2. 3. (w/ ) 4. • (Noisy) • • • Centroid
- 18. : • DAN Semantic Loss • Centroid •
- 20. • [ ] (Semantic Alignment ) • Semantic Alignment Noisy • Noise
- 21. 2. • Title: CyCADA: Cycle-Consistent Adversarial Domain Adaptation • Authors: Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, Trevor Darrell • Info: ICML2018 accepted (oral) • Abstract: (Pixel) Domain Translation CycleGAN
- 22. • • Semantic Alignment • • c.f. ( ) [Jo and Bengio, 2017] • ( ) • • • Cycle-GAN • • Cycle-GAN •
- 23. 4 Component • Cycle-GAN • Semantic Consistency • Classifier • Feature Adaptation
- 25. : Semantic Consistency Loss (f_s pre-train )
- 26. : Classifier
- 27. : Feature Adaptation Domain Adversarial Network
- 28. : • SVHN -> MNIST Pixel Adaptation Feature Adaptation • Pixel Feature
- 29. : Cycle Loss Semantic Loss • SVHN -> MNIST Semantic Alignment • Semantic Segmentation( ) Cycle Loss Cycle
- 31. • • Cycle-GAN Cycle Loss Semantic Loss
- 32. • • • • [Saito+ 2018b]: • [Hosseini-Asl+ 2018]:
- 33. References • Y.Ganin,E.Ustinova,H.Ajakan,P.Germain,H.Larochelle, F. Laviolette, M. Marchand, and V. Lempitsky. Domain- adversarial training of neural networks. JMLR, 17(59):1–35, 2016. • Tzeng, E., Hoffman, J., Saenko, K., and Darrell, T. Adversarial discriminative domain adaptation. In Computer Vision and Pattern Recognition (CVPR), 2017. • Kuniaki Saito, Yoshitaka Ushiku, Tatsuya Harada. Asymmetric Tri-training for Unsupervised Domain Adaptation. The 34th International Conference on Machine Learning (ICML 2017), 2017. • Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, Tatsuya Harada. Maximum Classifier Discrepancy for Unsupervised Domain Adaptation. The 31th IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2018), 2018, (oral). • Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, Tatsuya Harada, Open Set Domain Adaptation by Backpropagation, arXiv, 2018b • Taigman, Y., Polyak, A., and Wolf, L. Unsupervised cross-domain image generation. In International Conference on Learning Representations, 2017a. • Shrivastava, A., Pfister, T., Tuzel, O., Susskind, J., Wang, W., and Webb, R. Learning from simulated and unsupervised im- ages through adversarial training. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. • Ehsan Hosseini-Asl, Yingbo Zhou, Caiming Xiong, Richard Socher, Augmented Cyclic Adversarial Learning for Domain Adaptation, arxiv, 2018 • Jason Jo, Yoshua Bengio, Measuring the tendency of CNNs to Learn Surface Statistical Regularities, arxiv 2017, https://arxiv.org/abs/1711.11561