GDG-Shanghai 2017 TensorFlow Summit Recap
8 likes•2,202 views
The document highlights the accomplishments of Jeff Dean, a key figure at Google known for his contributions to systems like MapReduce and TensorFlow. It discusses the evolution of machine learning, particularly through large-scale distributed systems and the challenges faced by previous models. Additionally, it covers the applications and future directions of TensorFlow in various fields, including healthcare and art generation.
GDG-Shanghai 2017 TensorFlow Summit Recap
- 1. 江骏 @ 饿了了么
- 2. Jeff Dean 的传说 source: http://blog.jobbole.com/51607/ • “Jeff Dean 是直接写⼆二进制机器器代码的。” • “编译器器从不不会给 Jeff Dean 警告的，Jeff Dean 会给编译器器警告的。” • “在 2000 年年末的时候，Jeff Dean 写代码的速度突然增⻓长了了 40 倍， 原因是他把⾃自⼰己的键盘升级到了了 USB 2.0。”
- 3. Jeff Dean • He has co-designed/implemented five generations of Google's • crawling • indexing • query serving systems • and co-designed/implemented major pieces of Google’s • initial advertising • AdSense for Content systems • Google's distributed computing infrastructure, including • MapReduce, BigTable, Spanner systems • protocol buffers, LevelDB, systems infrastructure for statistical machine translation • a variety of internal and external libraries and developer tools • He is currently working on large-scale distributed systems for machine learning. Leads the Google Brain team, Google's deep learning research team
- 4. • I’m a developer, not a data scientist. • Focus on playing TensorFlow in cloud. 江骏 @ 饿了了么
- 5. Keynote Video (Jeff Dean - part)
- 6. TensorFlow 的上⼀一代产品 DistBelief • 解决了了数据量量 > GPU 最⼤大内存的问题 • paper 中提到最⼤大的 model，达到了了 1.7 billion parameters, utilizing 81 machines, delivering a 12x speedup. • 它的缺点： • 擅⻓长图像识别，但对其它机器器学习 model 适⽤用性差，⽽而且不不⽀支持 mobile • 维护⼤大规模系统成为负担，缺乏抽象。 https://research.google.com/pubs/pub40565.html
- 11. Training Data Validation Data Test Data Model Serving request train (cpu/gpu) ✔ ✘
- 13. What is machine learning? Loss Weight Gradient Descent Bias Activation Function Hidden Layer Learning Rate
- 14. • 输⼊入 • 计算 ，把结果代⼊入 activation function，得到输出值wT x + b y x 神经元 neural • 预测得准吗？Loss function 来衡量量。⽐比如，⽤用它们的距离 • ⽬目标：让 loss 值越⼩小越好 • ⽤用 Optimizer 调整 weight 和 bias （⽐比如⽤用 Gradient Descent） loss = (y − y)2
- 15. Deep Learning
- 16. Python Golang Java … Graph (DAG) C++ Core compile run How does TensorFlow work
- 17. TensorFlow High-Level APIs & Keras & ML Toolkit Francois CholletMartin Wicke Ashish Agarwal
- 19. Low-Level APIs
- 20. High-Level APIs
- 26. Keras
- 27. Keras
- 29. Video Input
- 30. Question Input
- 31. Output
- 32. Generate the “model”, a graph
- 34. Wide & Deep Learning Heng-Tze Cheng https://www.tensorflow.org/tutorials/wide_and_deep/ research paper
- 40. Distributed TensorFlow TensorFlow Ecosystem: Integrating TensorFlow with Your Infrastructure Derek Murray Jonathan Hseu
- 48. TensorBoard Demo
- 49. Serving Models in Production with TensorFlow Serving Noah Fiedel https://github.com/tensorflow/serving
- 52. XLA (Accelerated Linear Algebra) Chris Leary and Todd Wang https://www.tensorflow.org/versions/master/experimental/xla/
- 59. TensorFlow at DeepMind Daniel Visentin Brett Kuprel Skin Cancer Image Classification Mobile and Embedded TensorFlow Pete Warden TensorFlow in Medicine Retinal Imaging Doug Eck Magenta: Music and Art Generation Eugene Brevdo Sequence Models and the RNN API Lily Peng