Easy Recipes to Make Apps using ML and Big Data

Presented by Mahsa
Created by Lucy

Transcript

1. 1 Easy Recipes to Make Apps using Machine Learning and Big Data Mahsa Orang Presented for Hack the North September 15, 2017
2. 2 Slides available on Slack channel #veeva Stop by our booth for questions and cool swag! We are hiring!
3. 3 About Me Mahsa Orang Associate Software Engineer at Veeva Machine Learning App for Similarity Search and Record Linkage Information Technology PhD BSc MSc Mathematics Computer Science
4. 4 Pic2Wiki App Working Example Database Pic2Wiki App DL Model Recommender  
5. 5 Agenda Today’s Recipes ML development environment Make your data ready for ML Use ready-to-use DL models Pre-process your big data Get results faster A bonus app Recipe 1 Recipe 2 Recipe 3 Recipe 4 Recipe 5 Recipe 6
6. 6 Background https://blogs.nvidia.com/wp-content/uploads/2016/07/Deep_Learning_Icons_R5_PNG.jpg.png Artificial Intelligence Machine Learning Deep Learning
7. 7 ML development environment Recipe 1 1 Make your data ready for ML Use ready-to-use DL models Pre-process your big data Get results faster A bonus app Recipe 2 Recipe 3 Recipe 4 Recipe 5 Recipe 6
8. 8 Set Up Your Notebook Apache Zeppelin Noteboook Jupyter Notebook https://zeppelin.apache.org/download.html http://jupyter.org/install.html
9. 9 What is Notebook?
10. 10 What is Notebook?
11. 11 What is Notebook?
12. 12 What is Notebook?
13. 13 ML development environment Recipe 1 Make your data ready for ML Use ready-to-use DL models Pre-process your big data Get results faster A bonus app Recipe 2 Recipe 3 Recipe 4 Recipe 5 Recipe 6 2
14. 14 Pic2Wiki App Working Example Database Pic2Wiki App DL Model Recommender  
15. 15 ML 101 Which one is the most similar to my pizza? My Pizza A B C
16. 16 ML 101 Which one is the most similar to my pizza? My Pizza 5 9 7 4 Convert data to real numbers
17. 17 ML 101 Which one is the most similar to my pizza? My Pizza 5 9 7 4 Then compare numbers 9-5=4 5-4=1 7-5=2 THIS ONE
18. 18 ML 101 Which one is the most similar to my pizza? Extract more real numbers <6,1,...> <5,2,...>
19. 19  ML 101 Which one is the most similar to my pizza? Calculate similarity between vectors <9,3,...> <5,2,...> <5,1,...> <6,1,...>
20. 20 ML 101 Calculate similarity between vectors 5 10 15 10 5 B Weight Diameter A
21. 21 ML 101 Calculate similarity between vectors 5 10 15 10 5 A B Weight Diameter Cosine Similarity
22. 22  ML 101 Which one is the most similar to my pizza? Calculate similarity between vectors <5,2,...> <9,1,...> <5,1,...> <6,1,...> 0.2 0.8 0.5 database input data
23. 23  ML 101 Which one is the most similar to my pizza? Calculate similarity between vectors <5,2,...> <9,1,...> <5,1,...> <6,1,...> 0.2 0.8 0.5 database input data
24. 24 ML development environment Recipe 1 Make your data ready for ML Use ready-to-use DL models Pre-process your big data Get results faster A bonus app Recipe 2 Recipe 3 Recipe 4 Recipe 5 Recipe 6 3
25. 25 Pic2Wiki App Working Example Database Pic2Wiki App DL Model Recommender  
26. 26  ML 101 Calculate distance between vectors <5,2,..> <9,3,...> <5,1,...> <6,1,...> 0.04 0.8 0.02 database input data
27. 27 Similar Objects are Nearby <5,1,...> <5,2,..> <9,3,...>
28. 28 ML/DL Pre-Trained Models 0 0.5 0 0.2 0.1 0
29. 29 ML/DL Pre-Trained Models 0 0.5 0 0.2 0.1 0
30. 30 ML/DL Pre-Trained Models 0 0.5 0 0.2 0.1 0 Inception http://download.tensorflow.org/models/inception_v4_2016_09_09.tar.gz
31. 31 ML/DL Pre-Trained Models 0 0.5 0 0.2 0.1 0 INCEPTION http://download.tensorflow.org/models/incep- tion_v4_2016_09_09.tar.gz WORD2VEC 0.9 0.5 0 0.7 0.1 1 https://drive.google.com/file/d/0B7XkCwpI5KDYNl- NUTTlSS21pQmM/edit?usp=sharing
32. 32 Get Feature Vector
33. 33 ML development environment Recipe 1 Make your data ready for ML Use ready-to-use DL models Pre-process your big data Get results faster A bonus app Recipe 2 Recipe 3 Recipe 4 Recipe 5 Recipe 6 4
34. 34 Pic2Wiki App Working Example Database Pic2Wiki App DL Model Recommender  
35. 35  Pre-process Your Big Data big data input data
36. 36  Pre-process Your Big Data big data input data 0.9 0.5 0 0.7 0.1 1 0.3 0.1 0.3 0.2 0.1 0 0 0.4 0 0 0.1 1 1 0.2 0.5 0.1 0.1 1
37. 37   Pre-process Your Big Data big data Loading Stage Vectorizing Stage    ML Worker ML Worker ML Worker
38. 38 Pre-process Your Big Data
39. 39 Pre-process Your Big Data 4 times faster
40. 40 ML development environment Recipe 1 Make your data ready for ML Use ready-to-use DL models Pre-process your big data Get results faster A bonus app Recipe 2 Recipe 3 Recipe 4 Recipe 5 Recipe 6 5
41. 41 How to Get Results Faster?  Input vector should be compared with all vectors! big data input data
42. 42  A Working Example Which nearby restaurant serves this drink?
43. 43 You are here A Working Example
44. 44 You are here A Working Example
45. 45 Longitude: 43.473 Latitude: -80.542 A Working Example
46. 46 GeoHash GeoHash https://anaconda.org/auto/python-geohash Longitude Latitude
47. 47 Longitude: 43.473 Latitude: -80.542 Geohash: dpwxr1 A Working Example Longitude: 43.472 Latitude: -80.544 Geohash: dpwxr1 UW Math Coffee and Donuts
48. 48 1. Find all geohashes offline dpwxr1
49. 49 2. Find input geohash dpwxr1 Loading nearby restaurants ... Search Input Geohash: dpwxr1
50. 50 ML development environment Recipe 1 Make your data ready for ML Use ready-to-use DL models Process your Big Data offline Get results faster A bonus app Recipe 2 Recipe 3 Recipe 4 Recipe 5 Recipe 6 6
51. 51 Build AI Apps Using Transfer Learning
52. 52 Deep Learning image (pixel) Input Layer Output Layer GEOFFREY HINTON Layers of Abstraction DEEP NEURAL NETWORK (DNN) object models edges object parts
53. 53 Transfer Learning Input Layer Output Layer Layers of Abstraction PRE-TRAINED DL MODEL Replace last layers CLASSIFIER Hotdog or Not Hotdog
54. 54 How to Build a Classifer in 5 min Not Hotdog App DL Model Classifier  input data output data Hotdog Not Hotdog or
55. 55 Not Hotdog App DL Model Classifier  input data output data Hotdog Not Hotdog or How to Build a Classifer in 5 min
56. 56 How to Build a Classifer in 5 min Not Hotdog App DL Model Classifier  input data output data Hotdog Not Hotdog or
57. 57  How to Build a Classifer in 5 min data
58. 58  0 0.4 0 0 0.1 1 How to Build a Classifer in 5 min 1 0.2 0.5 0.1 0.1 1 1. Find all the vectors of the images using DL model 0.3 0.1 0.3 0.2 0.1 0 data hotdog pizza pasta
59. 59  0 0.4 0 0 0.1 1 How to Build a Classifer in 5 min 1 0.2 0.5 0.1 0.1 1 2. Find the vector of the input image using DL model 0.3 0.1 0.3 0.2 0.1 0 data hotdog pizza pasta input data
60. 60  0 0.4 0 0 0.1 1 How to Build a Classifer in 5 min 1 0.2 0.5 0.1 0.1 1 2. Find the vector of the input image using DL model 0.3 0.1 0.3 0.2 0.1 0 data hotdog pizza pasta input data 0.8 0.7 0.2 0.2 0.1 0
61. 61  0.82 0.1 0.02 0 0.4 0 0 0.1 1 How to Build a Classifer in 5 min 1 0.2 0.5 0.1 0.1 1 3. Compare all the vectors using cosine similarity 0.3 0.1 0.3 0.2 0.1 0 data hotdog pizza pasta input data 0.8 0.7 0.2 0.2 0.1 0
62. 62  0.82 0.1 0.02 0 0.4 0 0 0.1 1 How to Build a Classifer in 5 min 1 0.2 0.5 0.1 0.1 1 4. Find the most similar class 0.3 0.1 0.3 0.2 0.1 0 data hotdog pizza pasta input data 0.8 0.7 0.2 0.2 0.1 0
63. 63  0.82 0.1 0.02 0 0.4 0 0 0.1 1 1 0.2 0.5 0.1 0.1 1 0.3 0.1 0.3 0.2 0.1 0 data hotdog pizza pasta input data 0.8 0.7 0.2 0.2 0.1 0 How to Build a Classifer in 5 min 4. Find the most similar class
64. 64  0.82 0.1 0.02 0 0.4 0 0 0.1 1 1 0.2 0.5 0.1 0.1 1 0.3 0.1 0.3 0.2 0.1 0 data hotdog pizza pasta input data 0.8 0.7 0.2 0.2 0.1 0 How to Build a Classifer in 5 min 4. Find the most similar class hotdog!
65. 65  0 0.4 0 0 0.1 1 Let’s Try Another One! 1 0.2 0.5 0.1 0.1 1 1. Find all the vectors of the images using DL model 0.3 0.1 0.3 0.2 0.1 0 data hotdog pizza pasta input data
66. 66  0 0.4 0 0 0.1 1 Let’s Try Another One! 1 0.2 0.5 0.1 0.1 1 2. Find the vector of the input image using DL model 0.3 0.1 0.3 0.2 0.1 0 data hotdog pizza pasta input data 1 0.7 0.2 0.2 0.1 0.9
67. 67  0.2 0.1 0.7 0 0.4 0 0 0.1 1 Let’s Try Another One! 1 0.2 0.5 0.1 0.1 1 3. Compare all the vectors using cosine similarity 0.3 0.1 0.3 0.2 0.1 0 data hotdog pizza pasta input data 1 0.7 0.2 0.2 0.1 0.9
68. 68  0.2 0.1 0.7 0 0.4 0 0 0.1 1 Let’s Try Another One! 1 0.2 0.5 0.1 0.1 1 3. Compare all the vectors using cosine similarity 0.3 0.1 0.3 0.2 0.1 0 data hotdog pizza pasta input data 1 0.7 0.2 0.2 0.1 0.9 NOT hotdog!
69. 69 1NN Classifier  Hotdog Not Hotdog
70. 70 Transfer Learning Input Layer Output Layer Layers of Abstraction PRE-TRAINED DL MODEL Replace last layers 1NN Classifier  Hotdog Not Hotdog
71. 71 Transfer Learning Input Layer (pixels) Prediction Layers of Abstraction NOT HOTDOG APP 1NN Classifier  Hotdog Not Hotdog Input Image Hotdog Not Hotdog or
72. 72 Summary Recipe 1: Set up your notebook Apache Zeppelin Noteboook Jupyter Notebook https://zeppelin.apache.org/download.html http://jupyter.org/install.html
73. 73 Summary Recipe 2: Make your data ready for ML <6,1,...> <5,2,...> Weight 5 10 15 10 5 A B Weight Diameter Cosine Similarity
74. 74 Summary Recipe 3: Use ready-to-use DL models 0 0.5 0 0.2 0.1 0 INCEPTION WORD2VEC 0.9 0.5 0 0.7 0.1 1
75. 75 Summary Recipe 4: Process your Big Data offline   big data Loading Stage Vectorizing Stage    ML Worker ML Worker ML Worker
76. 76 Summary Recipe 5: Get results faster dpwxr1 Loading nearby restaurants ... Search Input Geohash: dpwxr1
77. 77 Summary Recipe 6: A bonus app
78. 78 Slides available on Slack channel #veeva Stop by our booth for questions and cool swag! We are hiring!
79. 79 Our Team Nicole Employee Success Lucy Interaction Designer Iman Software Engineer Caleb Director of PM Amanda Director of UX Mahsa Software Engineer Mark Director of Eng
80. 80  Questions