1 / 22

ECE 599/692 – Deep Learning Lecture 1 1 – Generative Adversarial Networks

ECE 599/692 – Deep Learning Lecture 1 1 – Generative Adversarial Networks. Hairong Qi, Gonzalez Family Professor Electrical Engineering and Computer Science University of Tennessee, Knoxville http://www.eecs.utk.edu/faculty/qi Email: hqi@utk.edu. Outline.

jjane
Download Presentation

ECE 599/692 – Deep Learning Lecture 1 1 – Generative Adversarial Networks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. ECE 599/692 – Deep LearningLecture 11 – GenerativeAdversarialNetworks Hairong Qi, Gonzalez Family Professor Electrical Engineering and Computer Science University of Tennessee, Knoxville http://www.eecs.utk.edu/faculty/qi Email: hqi@utk.edu

  2. Outline Lecture 11: GAN–Introductionandtheoreticanalysis Lecture 12: Theoreticanalysisandothergenerativemodels–VAE Lecture 13: ConditionalGAN Lecture14:Implementation

  3. GAN • Twoneuralnetworkscompeteagainsteachother • AgeneratornetworkG:mimictrainingsamplestofoolthediscriminator • AdiscriminatornetworkD:discriminatetrainingsamplesandgeneratedsamples

  4. GAN

  5. GAN-Drawbacks Modemissingproblem Generateunrealisticimages Hardtolearntogeneratediscretedata,e.g.,text

  6. Evolution of GAN

  7. Image In-painting [13,14] [13] Pathak, Deepak, et al. "Context encoders: Feature learning by inpainting." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016. [14] Li, Yijun, et al. "Generative Face Completion." CVPR (2017).

  8. Image Editing [11] Larsen, Anders Boesen Lindbo, et al. "Autoencoding beyond pixels using a learned similarity metric.”ICML(2016).

  9. Image Blending [16] [16]Wu, Huikai, et al. "GP-GAN: Towards Realistic High-Resolution Image Blending." arXiv preprint arXiv:1703.07195 (2017).

  10. AgeProgressionandRegression Continuously bidirectional aging Input Others Ours 0~5 6~10 11~15 16~20 21~30 31~40 41~50 51~60 61~70 71~80 8 31~40 31~40 5 69~80 71~80 7 16~20 16~20 45 60~80 61~70 Project page: https://zzutk.github.io/Face-Aging-CAAE

  11. GAN–TheoreticReasoning

  12. Theoretic Reasoning • Instability of GAN • Mode missing • Gradient vanishing

  13. Instability – Mode Missing Expectation: Usually, G and D are updated alternatively

  14. Instability – Mode missing (cont’d)

  15. Instability – Mode missing (cont’d)

  16. Instability – Gradient vanishing

  17. Instability Don’t train “D” too good or too poor! If “D” is optimal, it will cause gradient vanishing for G. If “D” is poor, the gradient of G is unstable, huge occlusion.

  18. GAN-based Image Manipulation

  19. Stabilizing GAN –Adding an AE In GAN, the generated distribution, pg(x), is matched to the distribution specified by D, rather than the real distribution

  20. Stabilizing GAN - adding an AE (cont’d) Remedy to Unrealistic Generation

  21. Acknowledgement Most slides are taken from student presentations at AICIP group meetings, including mainly those from Zhifei Zhang.

More Related