ºÝºÝߣ

ºÝºÝߣShare a Scribd company logo

[????] ??? 2018 ??

?
?
D
Donghyeon Kim

* ???????? 2018? 12?? ??? Moducon 2018? ?????. * ??? ??? ?? ? ??? ????? 1. Research of Clova AI toward 'AI for Everyone' - ??? ? (Clova AI Research Director) 2. ?? ???? ?? - ??? ? (?????) * ??????? ??????? A-GIST ???? ??????. * ???? (???, ???): https://youtu.be/FRvlwaqrGHM

1 of 40
Downloaded 16 times
MODUCON 2018 Review 2019. 02. 26. ??? dhkim518@gist.ac.kr
??? ??? 2 http://www.modulabs.co.kr/ https://www.facebook.com/lab4all/ https://github.com/modulabs
MODUCON 2018 3 ? ?????? ? ???? (2018. 12. 15.) ? AI ???? ??, Deep Learning College(DLC) 1? ????, ?? ? ???? Tutorial Session ? ?? ??? ??, ?? ??/ ?? ?? ?? http://moducon.kr/2018/
?? ??? ?? 4 ¡ù ? ?? ??? MODUCON 2018 ? ??? ??? ???? ??????
5
?? ? ??? ????? NAVER ? LINE 6
? Clova: Cloud-based Virtual Assistant Clova: AI Assistant Platform of NAVER & LINE 7 Vision Speech Recognition/ Synthesis Recommen- dation Natural Language Understanding Machine Learning Deep Learning Model Research
? OCR: Optical Character Recognition ?????? (1) Detection: ????? text ???? https://github.com/hwalsuklee/awesome-deep-text-detection-recognition?fbclid=IwAR2844IsN0PG_G0Wwt_PkOBQTVhjCv3xbptEDlPiyzp6u7pFOD9m9qDTS-8 1. OCR 8 Remarks - CVPR2019 - Curved text ?? ??
? OCR: Optical Character Recognition ?????? (2) Recognition: ??? text ???? https://github.com/hwalsuklee/awesome-deep-text-detection-recognition?fbclid=IwAR2844IsN0PG_G0Wwt_PkOBQTVhjCv3xbptEDlPiyzp6u7pFOD9m9qDTS-8 1. OCR 9 Remarks - OCR ??? ?? ??? ????, ??? ??? ?? ?? - ? ???? ?? ???? ?? ? ??, ?? ??? ?? ? ???? - ???? ??? ??? (??? ????? ?? ??? ?? ??)
? Applications of tracking and re-id in in-the-wild video ? AutoCAM (www.vlive.tv/search/all?query=autocam) 2. Video AI 10
? Applications of tracking and re-id in in-the-wild video ? AutoCAM (www.vlive.tv/search/all?query=autocam) 2. Video AI 11
? Applications of tracking and re-id in in-the-wild video ? AutoCAM (www.vlive.tv/search/all?query=autocam) 2. Video AI 12 ? ??? ?? ??? ????? ?? ??? ???? ? AutoCAM: ?? ? ?????? ?? ??? ??? ?? ?? ??? ? ??? ??? ?? Tracking + Personal Re-identification + Pose Estimation + Face Recognition ? ?? ??? ???? ? 30 % ? Clova ??? 90 % ?? ??? ?? ? Challenging Issues ?? ????, ??? ????, ?? ??? ???¡­
? Applications of tracking and re-id in in-the-wild video ? AutoCut (www.vlive.tv/search/all?query=autocut) 2. Video AI 13
? Applications of tracking and re-id in in-the-wild video ? AutoCut (www.vlive.tv/search/all?query=autocut) 2. Video AI 14 ? ??? ?? ??? ????? ?? ??? ???? ? ??,???????? 1?? ?? ?? ? ? ??? 10? ?? ??? ? AutoCut: ?? ?? ?????? ?? ??? ??? ??? ????
? Wavelet Pooling for Perfect Reconstruction ? Max pooling? perfect reconstruction? ? ? ?? ??? ? Wavelet pooling? ??? ??? ???? ?? 3. Super-Real Style Transfer 15 Yoo et al. 2018
? Wavelet Pooling for Perfect Reconstruction ? Max pooling? perfect reconstruction? ? ? ?? ??? ? Wavelet pooling? ??? ??? ???? ?? 3. Super-Real Style Transfer 16 Clova AI: ?? ???
? Super-real Style Transfer for Video ? Photo-realistic ? ?? ??? ?????? ???: Flickering? ??? ??? ?? ??? 3. Super-Real Style Transfer 17
18 ¡ù ? ??? ??? ??? edwith ?? ??? ?? ???? ??????. https://www.edwith.org/bayesiandeeplearning/lecture/25292/
? ???? ??/?? Motivation 19 ??? ??, edwith ?? ??, https://www.edwith.org/bayesiandeeplearning/lecture/25292/
? Uncertainty: ??? ? ???? ??? ??? ???? Motivation 20 MNIST(0~9)? ??? ??? ??
? Uncertainty: ??? ? ???? ??? ??? ???? Motivation 21 MNIST(0~9)? ??? ??? ?? Input?? C ? ???? ???..
Paper 22 2016.12.05. arXiv 2017 NIPS
??? ?? 23 ??? ?, MODUCON 2018 ????
Non-Bayesian ?? 24 ??? ??, edwith ?? ??, https://www.edwith.org/bayesiandeeplearning/lecture/25292/
Non-Bayesian ?? 25 ??? ??, edwith ?? ??, https://www.edwith.org/bayesiandeeplearning/lecture/25292/
? Bayesian Neural Network (BNN) ? ??? Uncertainty ??? ?? BNN ? ???? ???? ? ???? Neural Network? ?? Training ??? ?? ??? ???? ? ???? Neural Network? ?? ?? ???? ???? ? ? ???? ???? ?? ? Non-Bayesian ?? ? Ensemble ? ???? ?? ??? ????? ? Neural Network ??? ??? ????? ??? ?? ?? ??? ?? ? ??? Uncertainty ?? ??? ?? ? Remark ? ????? ??? ???? ?????? ???? ??? ???? Uncertainty ??? ?? ??? ??? ? ?? Non-Bayesian ?? 26 ??? ?, MODUCON 2018 ????
? Ensemble Learning Method ? ??? ?? ????? ?? ???? ?? ? ? ????? ????? ???? ?? ??? ???? Ensemble Model 27 ??? ?, MODUCON 2018 ????
? Ensemble Learning Method ? ??? ?? ????? ?? ???? ?? ? ? ????? ????? ???? ?? ??? ???? Ensemble Model 28 Remark Input X? ??? Output Y ??? ?? ?? ???, Y? ????? ???: ??? ?? ??? ?, MODUCON 2018 ????
? Classification ? Brier Score (BS): ??? ????? ???? Loss function ? ?? ? ?? Label ? one-hot ??? network? output ??? squared error ? ?? ? Regression ? ???? regression ???? NN? single value output ? ?? ? ?? ???? ?? ?? Mean Squared Error (MSE)? ?? ?? ? ??? MSE? uncertainty? ???? ??? ? Negative Log-likelihood (NLL) Ensemble Model 29 ??? ?, MODUCON 2018 ????
(Optional ? ????, ? ??? ??? ?? ??) ? Remark. Data augmentation? ?? ?? ? Explaining and Harnessing Adversarial Examples ?? ?? ?? ? I.J. Goodfellow, J. Shlens, and C. Szegedy in ICLR, 2015 ? Fast Gradient Sign Method? ???? adversarial examples ?? ? Adversarial perturbation? network? loss ? ???? ???? training data? perturbation? ??? ? Classifier? Robustness ?? Adversarial Training 30 ???????? ??? ?, MODUCON 2018 ????
2. ? ????? Parameters Initialization 3. M?? ????? ?? ?? 4. ?? ?????? ? ????? ????? ?? Mini batch dataset ?? 5. Adversarial Example ?? (Optional) 6. Loss? ??? ??? ? ????? ???? ?? ?? ?? 31 ??? ?, MODUCON 2018 ????
? MNIST & Not-MNIST Result: Classification 32 ??? ?, MODUCON 2018 ????
? MC-Dropout ? ?? ?? ? Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning ? R: Random signed vector ? Adversarial example? ?? ??? ?? ?? ? AT: Adversarial Training Result: Classification 33 MNIST? ?? MNIST? ??? MNIST? ?? Not-MNIST? ??? *???? 0: ????? ?? *Ensemble? ?? ???? ?? ??? *MC dropout ?? Ensemble ??? ?? ??? ?, MODUCON 2018 ????
? Known: Dogs (ImageNet), Unknown: Non-dogs (ImageNet) Result: Classification 34 Dogs? ?? Dogs? ??? Dogs? ?? Non-Dogs? ??? Dogs? ?? Dogs? ??? Dogs? ?? Non-Dogs? ??? Entropy Values *Dogs-Dogs ? ???? 0: ????? ?? *Dogs-Ndogs ? Ensemble ??? ????? ????? ?? ??? Max. Predicted Probability *Dogs-Dogs ? ?? ??? ?? 1?? *Dogs-Ndogs ? Ensemble ??? ????? ?? ??? ?? ???? ??? ?, MODUCON 2018 ????
Result: Regression 35 Empirical Variance (5) using MSE Density Network (1) using NLL Adversarial Training Deep Ensemble (5) Density Network Trick Positive ??? ????? ?? Exp() ??? ?? Empirical Variance (5) Deep Ensemble (5) Ground truth curve Observed noisy training data Predicted mean along with standard deviation ??? ??, edwith ?? ??, https://www.edwith.org/bayesiandeeplearning/lecture/25292/
? https://github.com/Kyushik/Predictive-Uncertainty-Estimation-using-Deep-Ensemble Further Info. (???? Github) 36 ??? ?, Github ??, https://github.com/Kyushik/Predictive-Uncertainty-Estimation-using-Deep-Ensemble
? https://github.com/Kyushik/Predictive-Uncertainty-Estimation-using-Deep-Ensemble Further Info. (???? Github) 37 ??? ?, Github ??, https://github.com/Kyushik/Predictive-Uncertainty-Estimation-using-Deep-Ensemble
? ?? ?: ???? ?? ? ???? ??? ?? ???? ?? ? ????? ???? ??? ??? ???? ?? ? ?? ??? ???? ?? ? ? ??? ? ??! Performance? ??! ? ???? ???? ? (??? ??? ???? ??? ??? ? ???) ? ?? ?? ?? ?? ??/ ?? ?? ??? ?? ?????.. ? ??? ??? ??? ??? ??: ? ?? ???? ???, ??? ???? ???.. ? ??????? ???????? ?? ???? ???? ? 2019 MODUCON ? ?????:) MODUCON 2018 ?? 38
Thank you 39
? MC-Dropout ? ?? ?? ? Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning ? R: Random signed vector ? Adversarial example? ?? ??? ?? ?? ? AT: Adversarial Training ? AT? ???? Ensemble ???? ??? ????? ? Classification ???? ???, Loss (score) ?? NLL? BS ?? ?? Appendix. Result: Classification 40

More Related Content

[????] ??? 2018 ??

  • 1. MODUCON 2018 Review 2019. 02. 26. ??? dhkim518@gist.ac.kr
  • 3. MODUCON 2018 3 ? ?????? ? ???? (2018. 12. 15.) ? AI ???? ??, Deep Learning College(DLC) 1? ????, ?? ? ???? Tutorial Session ? ?? ??? ??, ?? ??/ ?? ?? ?? http://moducon.kr/2018/
  • 4. ?? ??? ?? 4 ¡ù ? ?? ??? MODUCON 2018 ? ??? ??? ???? ??????
  • 5. 5
  • 6. ?? ? ??? ????? NAVER ? LINE 6
  • 7. ? Clova: Cloud-based Virtual Assistant Clova: AI Assistant Platform of NAVER & LINE 7 Vision Speech Recognition/ Synthesis Recommen- dation Natural Language Understanding Machine Learning Deep Learning Model Research
  • 8. ? OCR: Optical Character Recognition ?????? (1) Detection: ????? text ???? https://github.com/hwalsuklee/awesome-deep-text-detection-recognition?fbclid=IwAR2844IsN0PG_G0Wwt_PkOBQTVhjCv3xbptEDlPiyzp6u7pFOD9m9qDTS-8 1. OCR 8 Remarks - CVPR2019 - Curved text ?? ??
  • 9. ? OCR: Optical Character Recognition ?????? (2) Recognition: ??? text ???? https://github.com/hwalsuklee/awesome-deep-text-detection-recognition?fbclid=IwAR2844IsN0PG_G0Wwt_PkOBQTVhjCv3xbptEDlPiyzp6u7pFOD9m9qDTS-8 1. OCR 9 Remarks - OCR ??? ?? ??? ????, ??? ??? ?? ?? - ? ???? ?? ???? ?? ? ??, ?? ??? ?? ? ???? - ???? ??? ??? (??? ????? ?? ??? ?? ??)
  • 10. ? Applications of tracking and re-id in in-the-wild video ? AutoCAM (www.vlive.tv/search/all?query=autocam) 2. Video AI 10
  • 11. ? Applications of tracking and re-id in in-the-wild video ? AutoCAM (www.vlive.tv/search/all?query=autocam) 2. Video AI 11
  • 12. ? Applications of tracking and re-id in in-the-wild video ? AutoCAM (www.vlive.tv/search/all?query=autocam) 2. Video AI 12 ? ??? ?? ??? ????? ?? ??? ???? ? AutoCAM: ?? ? ?????? ?? ??? ??? ?? ?? ??? ? ??? ??? ?? Tracking + Personal Re-identification + Pose Estimation + Face Recognition ? ?? ??? ???? ? 30 % ? Clova ??? 90 % ?? ??? ?? ? Challenging Issues ?? ????, ??? ????, ?? ??? ???¡­
  • 13. ? Applications of tracking and re-id in in-the-wild video ? AutoCut (www.vlive.tv/search/all?query=autocut) 2. Video AI 13
  • 14. ? Applications of tracking and re-id in in-the-wild video ? AutoCut (www.vlive.tv/search/all?query=autocut) 2. Video AI 14 ? ??? ?? ??? ????? ?? ??? ???? ? ??,???????? 1?? ?? ?? ? ? ??? 10? ?? ??? ? AutoCut: ?? ?? ?????? ?? ??? ??? ??? ????
  • 15. ? Wavelet Pooling for Perfect Reconstruction ? Max pooling? perfect reconstruction? ? ? ?? ??? ? Wavelet pooling? ??? ??? ???? ?? 3. Super-Real Style Transfer 15 Yoo et al. 2018
  • 16. ? Wavelet Pooling for Perfect Reconstruction ? Max pooling? perfect reconstruction? ? ? ?? ??? ? Wavelet pooling? ??? ??? ???? ?? 3. Super-Real Style Transfer 16 Clova AI: ?? ???
  • 17. ? Super-real Style Transfer for Video ? Photo-realistic ? ?? ??? ?????? ???: Flickering? ??? ??? ?? ??? 3. Super-Real Style Transfer 17
  • 18. 18 ¡ù ? ??? ??? ??? edwith ?? ??? ?? ???? ??????. https://www.edwith.org/bayesiandeeplearning/lecture/25292/
  • 19. ? ???? ??/?? Motivation 19 ??? ??, edwith ?? ??, https://www.edwith.org/bayesiandeeplearning/lecture/25292/
  • 20. ? Uncertainty: ??? ? ???? ??? ??? ???? Motivation 20 MNIST(0~9)? ??? ??? ??
  • 21. ? Uncertainty: ??? ? ???? ??? ??? ???? Motivation 21 MNIST(0~9)? ??? ??? ?? Input?? C ? ???? ???..
  • 23. ??? ?? 23 ??? ?, MODUCON 2018 ????
  • 24. Non-Bayesian ?? 24 ??? ??, edwith ?? ??, https://www.edwith.org/bayesiandeeplearning/lecture/25292/
  • 25. Non-Bayesian ?? 25 ??? ??, edwith ?? ??, https://www.edwith.org/bayesiandeeplearning/lecture/25292/
  • 26. ? Bayesian Neural Network (BNN) ? ??? Uncertainty ??? ?? BNN ? ???? ???? ? ???? Neural Network? ?? Training ??? ?? ??? ???? ? ???? Neural Network? ?? ?? ???? ???? ? ? ???? ???? ?? ? Non-Bayesian ?? ? Ensemble ? ???? ?? ??? ????? ? Neural Network ??? ??? ????? ??? ?? ?? ??? ?? ? ??? Uncertainty ?? ??? ?? ? Remark ? ????? ??? ???? ?????? ???? ??? ???? Uncertainty ??? ?? ??? ??? ? ?? Non-Bayesian ?? 26 ??? ?, MODUCON 2018 ????
  • 27. ? Ensemble Learning Method ? ??? ?? ????? ?? ???? ?? ? ? ????? ????? ???? ?? ??? ???? Ensemble Model 27 ??? ?, MODUCON 2018 ????
  • 28. ? Ensemble Learning Method ? ??? ?? ????? ?? ???? ?? ? ? ????? ????? ???? ?? ??? ???? Ensemble Model 28 Remark Input X? ??? Output Y ??? ?? ?? ???, Y? ????? ???: ??? ?? ??? ?, MODUCON 2018 ????
  • 29. ? Classification ? Brier Score (BS): ??? ????? ???? Loss function ? ?? ? ?? Label ? one-hot ??? network? output ??? squared error ? ?? ? Regression ? ???? regression ???? NN? single value output ? ?? ? ?? ???? ?? ?? Mean Squared Error (MSE)? ?? ?? ? ??? MSE? uncertainty? ???? ??? ? Negative Log-likelihood (NLL) Ensemble Model 29 ??? ?, MODUCON 2018 ????
  • 30. (Optional ? ????, ? ??? ??? ?? ??) ? Remark. Data augmentation? ?? ?? ? Explaining and Harnessing Adversarial Examples ?? ?? ?? ? I.J. Goodfellow, J. Shlens, and C. Szegedy in ICLR, 2015 ? Fast Gradient Sign Method? ???? adversarial examples ?? ? Adversarial perturbation? network? loss ? ???? ???? training data? perturbation? ??? ? Classifier? Robustness ?? Adversarial Training 30 ???????? ??? ?, MODUCON 2018 ????
  • 31. 2. ? ????? Parameters Initialization 3. M?? ????? ?? ?? 4. ?? ?????? ? ????? ????? ?? Mini batch dataset ?? 5. Adversarial Example ?? (Optional) 6. Loss? ??? ??? ? ????? ???? ?? ?? ?? 31 ??? ?, MODUCON 2018 ????
  • 32. ? MNIST & Not-MNIST Result: Classification 32 ??? ?, MODUCON 2018 ????
  • 33. ? MC-Dropout ? ?? ?? ? Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning ? R: Random signed vector ? Adversarial example? ?? ??? ?? ?? ? AT: Adversarial Training Result: Classification 33 MNIST? ?? MNIST? ??? MNIST? ?? Not-MNIST? ??? *???? 0: ????? ?? *Ensemble? ?? ???? ?? ??? *MC dropout ?? Ensemble ??? ?? ??? ?, MODUCON 2018 ????
  • 34. ? Known: Dogs (ImageNet), Unknown: Non-dogs (ImageNet) Result: Classification 34 Dogs? ?? Dogs? ??? Dogs? ?? Non-Dogs? ??? Dogs? ?? Dogs? ??? Dogs? ?? Non-Dogs? ??? Entropy Values *Dogs-Dogs ? ???? 0: ????? ?? *Dogs-Ndogs ? Ensemble ??? ????? ????? ?? ??? Max. Predicted Probability *Dogs-Dogs ? ?? ??? ?? 1?? *Dogs-Ndogs ? Ensemble ??? ????? ?? ??? ?? ???? ??? ?, MODUCON 2018 ????
  • 35. Result: Regression 35 Empirical Variance (5) using MSE Density Network (1) using NLL Adversarial Training Deep Ensemble (5) Density Network Trick Positive ??? ????? ?? Exp() ??? ?? Empirical Variance (5) Deep Ensemble (5) Ground truth curve Observed noisy training data Predicted mean along with standard deviation ??? ??, edwith ?? ??, https://www.edwith.org/bayesiandeeplearning/lecture/25292/
  • 36. ? https://github.com/Kyushik/Predictive-Uncertainty-Estimation-using-Deep-Ensemble Further Info. (???? Github) 36 ??? ?, Github ??, https://github.com/Kyushik/Predictive-Uncertainty-Estimation-using-Deep-Ensemble
  • 37. ? https://github.com/Kyushik/Predictive-Uncertainty-Estimation-using-Deep-Ensemble Further Info. (???? Github) 37 ??? ?, Github ??, https://github.com/Kyushik/Predictive-Uncertainty-Estimation-using-Deep-Ensemble
  • 38. ? ?? ?: ???? ?? ? ???? ??? ?? ???? ?? ? ????? ???? ??? ??? ???? ?? ? ?? ??? ???? ?? ? ? ??? ? ??! Performance? ??! ? ???? ???? ? (??? ??? ???? ??? ??? ? ???) ? ?? ?? ?? ?? ??/ ?? ?? ??? ?? ?????.. ? ??? ??? ??? ??? ??: ? ?? ???? ???, ??? ???? ???.. ? ??????? ???????? ?? ???? ???? ? 2019 MODUCON ? ?????:) MODUCON 2018 ?? 38
  • 40. ? MC-Dropout ? ?? ?? ? Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning ? R: Random signed vector ? Adversarial example? ?? ??? ?? ?? ? AT: Adversarial Training ? AT? ???? Ensemble ???? ??? ????? ? Classification ???? ???, Loss (score) ?? NLL? BS ?? ?? Appendix. Result: Classification 40