Icann2018ppt final
•
0 likes•51 views
1) The document presents a new deep unsupervised domain adaptation method that uses graph matching and pseudo-label guided training. 2) It introduces a second-order matching term to capture structural correspondence between domains, in addition to a first-order term. 3) The method performs a two-stage training, where in the first stage it reduces domain discrepancy using graph matching, and in the second stage it exploits unlabeled target data with pseudo-labels to further refine the decision boundaries.
![5
Introduction
Domain Adaptation Methods
Non-Deep Methods Deep Methods
• Instance Re-weighting
[Dai et al. ICML’07]
• Parameter Adaptation
[Bruzzone et al. TPAMI’10]
• Feature Transformation
[Fernando et al. ICCV’13]
[Sun et al. AAAI’16]
• Discrepancy Based
[Long et al. ICML’15]
[Sun et al. ECCV’16]
• Adversarial Based
[Ganin et al. JMLR’16]
[Tzeng et al. CVPR’17]](https://image.slidesharecdn.com/icann2018pptfinal-181009172352/85/Icann2018ppt-final-5-320.jpg)
![6
Introduction
• Discrepancy Based Methods
Mostly global metrics. Minimizes statistics of data like covariance
[Sun et al. ECCV’16] or maximum mean discrepancy [Long et al. ICML’15]
• Local Method
Optimal Transport (Courty et al. TPAMI’17). Basically point-point Matching
Using first order information might be misleading. How ?](https://image.slidesharecdn.com/icann2018pptfinal-181009172352/85/Icann2018ppt-final-6-320.jpg)
![17
Experimental Results
Dataset Settings
tunable
Batch Size – 32 source & target samples
Input – Decaf features [ICML‘14]
Feature Extractor – [500,100] FC ReLU
Optimizer – Adam [ICLR’15]
Implementation – Tensorflow on
NVIDIA GeForce GPUDataset available at https://github.com/RockySJ/WDGRL](https://image.slidesharecdn.com/icann2018pptfinal-181009172352/85/Icann2018ppt-final-17-320.jpg)
![18
Experimental Results
Comparison with previous work
MMD [JMLR’12], DANN [JMLR’16], CORAL [ECCV’16], WDGRL [AAAI’18]
(Maximum Mean Discrepancy) (Domain Adversarial) (Correlation alignment) (Wasserstein Distance)](https://image.slidesharecdn.com/icann2018pptfinal-181009172352/85/Icann2018ppt-final-18-320.jpg)
Icann2018ppt final
- 1. 1 Click to edit Master title style Graph Matching and Pseudo-Label Guided Deep Unsupervised Domain Adaptation Debasmit Das C.S. George Lee Assistive Robotics Technology Laboratory School of Electrical and Computer Engineering Purdue University, West Lafayette, IN, USA Funding Source : National Science Foundation (IIS-1813935)
- 2. 2 Outline • INTRODUCTION • OUR METHOD • EXPERIMENTAL RESULTS
- 3. 3 Introduction Classical Machine Learning Setting Find where Training and Testing Samples from same Distribution !!
- 4. 4 Introduction Classifying a Dog and a Cat Training Samples Testing Samples Training and Testing Distribution different!! Domain Adaptation Required!!
- 5. 5 Introduction Domain Adaptation Methods Non-Deep Methods Deep Methods • Instance Re-weighting [Dai et al. ICML’07] • Parameter Adaptation [Bruzzone et al. TPAMI’10] • Feature Transformation [Fernando et al. ICCV’13] [Sun et al. AAAI’16] • Discrepancy Based [Long et al. ICML’15] [Sun et al. ECCV’16] • Adversarial Based [Ganin et al. JMLR’16] [Tzeng et al. CVPR’17]
- 6. 6 Introduction • Discrepancy Based Methods Mostly global metrics. Minimizes statistics of data like covariance [Sun et al. ECCV’16] or maximum mean discrepancy [Long et al. ICML’15] • Local Method Optimal Transport (Courty et al. TPAMI’17). Basically point-point Matching Using first order information might be misleading. How ?
- 7. 7 Second order information Our Method
- 8. 8 Representing Correspondences Correspondence Matrix 1 0 Matching Matrix Continuous Relaxation Our Method
- 9. 9 First-order Matching Continuous Relaxation Our Method Continuous Relaxation Second-order Matching
- 10. 10 Our Method Optimization With additional correction Add equality constraints as penalty
- 12. 12 Closed form solution of mapping Iterative solution Our Method
- 13. 13 Second stage training procedure • Domain Discrepancy is reduced • Need to exploit unlabeled data • Chose confident unlabeled samples based on a threshold • ‘Sharpen’ the probability of these confident samples Our Method
- 14. 14 Pseudo-label (PL) Loss Our Method Class 2Class 1 Class 3 Sample 1 Sample 2 Sample 3
- 15. 15 Second stage training procedure Total Loss Our Method
- 16. 16 Overall architecture Stage 1 training Stage 2 training Our Method
- 17. 17 Experimental Results Dataset Settings tunable Batch Size – 32 source & target samples Input – Decaf features [ICML‘14] Feature Extractor – [500,100] FC ReLU Optimizer – Adam [ICLR’15] Implementation – Tensorflow on NVIDIA GeForce GPUDataset available at https://github.com/RockySJ/WDGRL
- 18. 18 Experimental Results Comparison with previous work MMD [JMLR’12], DANN [JMLR’16], CORAL [ECCV’16], WDGRL [AAAI’18] (Maximum Mean Discrepancy) (Domain Adversarial) (Correlation alignment) (Wasserstein Distance)
- 19. 19 Experimental Results Parameter Sensitivity Second-order term over First-order term Graph-Matching Loss over Classification loss Classification loss over Pseudo-label loss
- 21. 21 Experimental Results Feature Visualization Without Adaptation With Graph Matching With Graph Matching & Pseudo-labeling
- 22. 22 • Second order matching term is important to match structure in data • Refining decision boundaries by self-training with unlabeled data is beneficial • Performance improvement on image recognition justifies the inclusion of the 2-stage training Conclusion Future Work Multiple source domain generalization