Real-World Machine Learning with TensorFlow and Cloud ML

In this session, attendees will receive a brief introduction to machine learning and neural networks. Then, we will discuss various ways to use machine learning, from pre-trained model APIs for vision and language, to custom model development tools. We’ll also discuss how to deploy models into production. We will see how these technologies are applied in real-world customer scenarios.

Transcript

2. Big Data & AI Conference Dallas, Texas June 27 – 29, 2019 www.BigDataAIconference.com
3. Real-World Machine Learning with TensorFlow and Cloud AI Karl Weinmeister (@kweinmeister)
4. Agenda Intro to Machine Learning ML APIs AutoML Cloud AI Platform 1 2 3 4 5 TensorFlow
6. Confidential & Proprietary ML is Everywhere at Google
7. 5 What is machine learning?
8. 6 Finding patterns in data
9. 7 Classification Recommendation Forecasting Game Strategy
10. 8 Classification Recommendation Forecasting Game Strategy
11. 9 C a t Classifcati on Dog
12. 1 Build a model
13. 11 Labeled Training Dataset Examples of cats Examples of dogs
14. 12 Training a Model Model Machine Learning Algorithm
15. 13 Making Predictions Model “Dog”
16. 1 Finding patterns in data
17. 1 Finding patterns in data *requires data
18. 1 No data?
19. 1 Use Google’s models
20. Use a pre-trained model to accomplish common ML tasks Cloud Vision Cloud Natural Language Cloud Speech-to- Text & Text-to- Speech Cloud Video Intelligence Cloud Translation
21. Vision API cloud.google.com/vis ion/
22. labels
23. emotion
24. ocr “LUNA PARK”
25. landmark detection
26. Natural Language API cloud.google.com/natural- language/
28. Cloud Natural Language
29. Cloud Natural Language Arts & Entertainment / Visual Art & Design Food & Drinks / Restaurants / Pizzerias Computers & Electronics
30. 2 Got data? Train a custom model
31. 2 Cloud AutoML
32. Your training data Generate predictions with a REST API AutoM L Trai n Deplo y Serv e What is Cloud AutoML?
33. Visio n Table s Vide o Translati on Langua ge
34. AutoML Vision
35. Healthy or Pneumonia?
36. AutoML Tables
37. AutoML Tables Historic ofers from marketplace.xyz ID G e o Domai n Posted on: Title Description Categ ory Bran d ... Price sold: 10 4 U S market A Feb 1, 2018 “Dark red…” “Try this soft…” [“A, B, …”] Nike ... $92 20 4 U S market B Jan 20, 2018 “Women’s… ” “Medium-size…” [“A, B, …”] Adida s ... $58 30 2 U S market A Jan 12, 2018 “Running…” “All-terrain…” [“A, B, …”] Asics ... $85 35 2 EU market B Feb 13, 2018 “Running…” “All-terrain…” [“A, B, …”] Puma ... ?
38. AutoML Tables Target column Historic ofers from marketplace.xyz ID G e o Domai n Posted on: Title Description Categ ory Bran d ... Price sold: 10 4 U S market A Feb 1, 2018 “Dark red…” “Try this soft…” [“A, B, …”] Nike ... $92 20 4 U S market B Jan 20, 2018 “Women’s… ” “Medium-size…” [“A, B, …”] Adida s ... $58 30 2 U S market A Jan 12, 2018 “Running…” “All-terrain…” [“A, B, …”] Asics ... $85 35 2 EU market B Feb 13, 2018 “Running…” “All-terrain…” [“A, B, …”] Puma ... ?
39. AutoML Natural Language Classification
40. Movie Genre Classification “After discovering a mysterious artifact buried beneath the lunar surface, mankind sets off on a quest to find its origins with help from intelligent supercomputer HAL 9000.” sci-fi
41. That’s all you need to know!
42. But.. .
43. What’s AutoML doing?
44. Deep Neural Network
45. Classifying photos of cats and dogs
47. rgb(89,133,204)
52. 5 What if you have a custom task that requires more control over your model type and its inputs?
53. 51 Why consider custom model development? Use of other approaches such as unsupervised or reinforcement learning More control over model architecture and confguration Advanced feature engineering that includes domain expertise https://pixabay.com/vectors/steering-wheel-steering-wheel-150137/ https://pixabay.com/illustrations/craftsmen-mechanic-helm-workers-1020156/
54. Cloud AI Platform Training & Prediction cloud.google.com/ai- platform Distributed Training with GPUs and TPUs Fast and scalable online/batch prediction Fully managed platform for machine learning
55. Publish your machine learning workflow in an orchestrated pipeline Build custom AI applications by easily “stitching” and connecting various components like LEGO. Publish end-to-end pipelines that can be reused by other developers without worrying about the underlying details
56. 5 Let's take a closer look at building a custom model with a neural network
57. 55 TensorFlo w An open source Deep Learning library >1,800 contributors worldwide Apache 2.0 license Released by Google in 2015 TensorFlow 2.0 Easier to learn and use For beginners and experts Alpha available today
58. 56 Classificatio n Example … 0 1 2 3 4 5 6 7 8 9 model = tf.keras.models.Sequential([ tf.keras.layers.Flatte n(), tf.keras.layers.Dense(10,
59. 57 Thinking in Sequences Dogs are my favorite animals, I love how they wag their
60. 58 Recurrent Neural Networks
61. What are we going to predict? # of crime reports last week How many crimes will be reported today?
62. 6 Let's build a custom model with TensorFlow on AI Platform
63. Thank you!