Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Machine Learning in Google I/O 19

395 views

Published on

올해 Google I/O에서는 구글의 머신러닝 및 딥러닝 분야에 대한 다양한 접근이 소개되었습니다. 이 발표에서는 Google I/O 2019에서 다룬 머신러닝 세션들을 크게 머신러닝 플랫폼, 머신러닝 클라우드 및 머신러닝 기반의 응용 서비스 확장으로 구분하고, 각각에 대하여 요약해 봅니다. 또한 현재의 발표를 바탕으로 이후의 방향성이 어떻게 될 것인지에 대하여 몇가지 예측을 해 봅니다.

이 슬라이드는 2019년 6월 Google I/O Extended 판교 및 서울에서 발표한 슬라이드입니다.

This talk covers the machine learning activities published during Google I/O.

Published in: Technology

Machine Learning in Google I/O 19

  1. 1. Machine Learning in Google I/O 19 Jeongkyu Shin Google I/O `19 Extended Seoul / Pangyo
  2. 2. • Lablup Inc. • • Backend.AI • • Google Developer Expert • ML / DL GDE, Sprint Master
  3. 3. My Google I/O `19 • 41 • 20 • 7 • 5 GDE/GDE
  4. 4. extended
  5. 5. extended
  6. 6. “ ” ? :
  7. 7. • : AI → • “ ” “ ” ? • “ ?” → • “ ?” → • “ ?” → , , OS, Stadia, Fuchusia… “ ” “ ” Google I/O `19
  8. 8. • 14 + ⍺ • • 3 TensorFlow DevSummit …
  9. 9. + • Coral, Edge TPU • TensorFlow Lite • TensorFlow.js + PWA • Swift for TensorFlow • TensorFlow 2.0 • TF-Agents • • / • Federated Learning • On-chip • Private Join and Compute • TPU Pods V3 • Cloud TPU Pods Beta • ML Kit + Firebase :
  10. 10. : • Coral, Edge TPU • TensorFlow Lite • TensorFlow.js + PWA • Swift for TensorFlow
  11. 11. : • TensorFlow 2.0 • TF-Agents • (Machine Learning Fairness) • / •
  12. 12. • Federated Learning • On-chip • Private Join and Compute
  13. 13. • Federated Learning • On-chip • Private Join and Compute • I/O … • 2019 6 19 . • !
  14. 14. + • TPU Pods V3 • Cloud TPU Pods Beta • ML Kit + Firebase
  15. 15. : Coral, Edge TPU / TensorFlow Lite / TensorFlow.js + PWA / Swift for TensorFlow
  16. 16. Title Coral: Edge TPU
  17. 17. Coral • Edge TPU • • USB • PCI-E, SOM • • Mendel OS (Debian Fork) • TF Lite Edge TPU • Python SDK
  18. 18. • I/O `19: SDK • • ops : • • Edge AI ASIC • Neural ComputeStick / Jetson Nano • RP4 + Add-on • ? Coral
  19. 19. • TF • XLA • TensorFlow 2.0 / TensorFlow.js • GPU / NPU ( ) TensorFlow Lite
  20. 20. • node.js • prebuilt / TF Hub (AI Hub) TensorFlow.js
  21. 21. • node.js • prebuilt / TF Hub (AI Hub) TensorFlow.js
  22. 22. • • Backend Frontend • C++ + Python + (TF Lite) + … ! • Swift: • LLVM • ( Python + ) • On-the-track: (O’reilly) • Swift for TensorFlow
  23. 23. : TensorFlow 2.0/ TF-Agents / (Machine Learning Fairness) / /
  24. 24. • (6 ) • Keras / Eager Execution • • tf.compat.v1 • TF Docs Sprint Seoul TensorFlow 2.0 http://bit.ly/2Y2VdR5
  25. 25. • Jupyter Notebook • TensorFlow / pybullet • • TensorFlow 1.14 2.0 • 2019 6 7 : RL (GFootball) • https://github.com/google-research/ football TF TF-Agents https://github.com/tensorflow/agents
  26. 26. • • - , , • “ ” (ML Fairness)
  27. 27. • • • (Transparency Framework)
  28. 28. • • / What-if
  29. 29. • TensorFlow Model Analysis • • • •
  30. 30. : (Federated Learning) / On-chip
  31. 31. Federated Learning • (Data island) + • + + • 2016 • 2019 : TensorFlow Federated • PySyft (for PyTorch) *Sources: [1] https://ai.googleblog.com/2017/04/federated-learning-collaborative.html [2] Federated Learning: Machine Learning on Decentralized Data (Google I/O 19)
  32. 32. • • • • GBoard • / Federated Learning extended
  33. 33. TensorFlow TensorFlow Federated • API • • • API • • *Source: Federated Learning: Machine Learning on Decentralized Data
  34. 34. • • Pixel • • 1/10 • : • • ( ) • On-chip
  35. 35. Title + :
  36. 36. + • TPU Pods V3 • Cloud TPU Pods Beta • ML Kit + Firebase extended
  37. 37. • ? → ! TPU Pods ! Cloud TPU Pods Beta • XLNet ( 6 20 ) • BERT • Cloud TPU Pod 2.5 ! • 2.5 2.9 . extended
  38. 38. • • → → • / Transfer Learning Firebase + + AutoML Vision Edge
  39. 39. + • Coral, Edge TPU • TensorFlow Lite • TensorFlow.js + PWA • Swift for TensorFlow • TensorFlow 2.0 • TF-Agents • • / • Federated Learning • On-chip • Private Join and Compute • TPU Pods V3 • Cloud TPU Pods Beta • ML Kit + Firebase
  40. 40. ! ? facebook/jeongkyu.shininureyes@gmail.com inureyesgithub/inureyes

×