The Machine Learning Intervention

This talk is an attempt to provide the crucial information needed when you start doing Machine Learning work. Johann du Toit (our speaker) had some *aha* moments throughout his failures, and this talk tries to condense those learnings into one set of slides.

While talking and uncovering these topics in a simple to digest way he'll be sprinkling learnings and demos/examples from his recent work.

Johann will cover a few topics, including:

-> What machine learning actually is, and is not; to level the playing the playing field.
-> The different types of models
-> How you would go about training a model with examples to each
-> How to run your awesome model

We hope the takeaway from this talk will be a structured mindset for those now newly approaching machine learning; be they developers/designers or product managers. You should also be able to walk away with the ability to talk ML and know what to expect when being faced with an ML project.

For the experts, you might at least get to know a few new tools :)

Expect real-time demos and some wackiness sprinkled in.

Transcript

1. THE MACHINE LEARNING INTERVENTION
2. Johann du Toit
3. Google Developer Expert for Cloud since 2014 Google Developers Launchpad Mentor
4. HydraCorp.
6. me@johanndutoit.net Email Website @signedness
7. This Talk
13. THANKS!
15. This Talk Attem pt #2
21. ~ 25,000 attendees
22. ~ 6 buildings
28. • Framing the problem • Training a model • Running your model The plan
29. • Framing the problem • Training a model • Running your model The plan
30. • Find a suitable algorithm for your use case • Figure out how you are going to express your data 2 steps to framing
31. • Find a suitable algorithm for your use case • Figure out how you are going to express your data 2 steps to framing
32. What is Machine Learning?
35. "People worry that computers will get too smart and take over the world, but the real problem is that they're too stupid and they've already taken over the world." Pedro Domingos
42. Machine Learning AI
43. Machine learning (ML) is a category of algorithm that allows software applications to become more accurate in predicting outcomes without being explicitly programmed. Merriam-Webster
48. Linear Regression Decision Tree Naive Bayes K-Means Random Forest Logistic Regression Support Vector Machine (SVM) k- Nearest Neighbours (kNN)
49. Linear regression Predicting real numbers Classification Divide up into groups Reinforcement Learning Learn by doing and failing K-Means Find Patterns Unsupervised Supervised Supervised Supervised
50. Ensemble
51. Specific.
52. • Find a suitable algorithm for your use case • Figure out how you are going to express your data 2 steps to framing
53. "Your app/service is either collecting data or missing opportunities" Me
55. [ 11, 223.44, 5125.5 ]
56. • Text • Images • Sound Representation
57. • Text • Images • Sound Representation
58. [ 11, 223.44, 5125.5 ]
59. “hello world” != a number …
60. “cat1.jpg”,”cat”
 “dog1.jpg”,”dog” “cat2.jpg”,”cat”
 “cat3.jpg”,”cat”
61. One Hot Encoding
62. “cat1.jpg”,”0”
 “dog1.jpg”,”1” “cat2.jpg”,”0”
 “cat3.jpg”,”0”
63. • Text • Images • Sound Representation
67. [ 11, 223.44, 5125.5 ]
68. • Text • Images • Sound Representation
71. “hey”
73. “hey”
74. [ 11, 223.44, 5125.5 ]
75. Example
78. Weather that night
79. Weather that night How much pizza was bought Average number that pitch Average percentage that pitch
80. Weather that night How much pizza was bought Average number that pitch Average percentage that pitch Average RSVP
81. Weather that night How much pizza was bought Average number that pitch Average percentage that pitch Average RSVP Time of Meetup Location of Meetup
82. Weather that night How much pizza was bought Average number that pitch Average percentage that pitch Average RSVP Time of Meetup Location of Meetup
84. How much data do I need ?
85. depends.
86. “You can always hire a better machine learning engineer, but they won’t be able to fix having no data” Me ;)
87. • Framing the problem • Training a model • Running your model The plan
88. • Datasets • Tools • Running Training
89. • Datasets • Tools • Running Training
90. Training Set Test Set Don’t touch this until the very end The actual training data used
92. Linear Regression Decision Tree Naive Bayes K-Means Random Forest Logistic Regression Support Vector Machine (SVM) k- Nearest Neighbours (kNN)
93. Ensemble
94. Training Set Test Set Don’t touch this until the very end The actual training data used Validation Set For cross validation and model selection
95. pre-process
96. • Datasets • Tools • Running Training
100. Vs
101. Single Process
107. Auto ML / BigQuery-ML
118. • Datasets • Tools • Running Training
119. Local Demo
121. To the terminal matey!
122. AutoML
130. Considerations
133. Train Regularly
134. • Framing the problem • Training a model • Running your model The plan
135. • Hosted • Local Options
136. • Hosted • Local Options
138. Demo
140. • Hosted • Local Options
142. Demo
143. in closing
145. A model's blind spots reflect the judgments and priorities of its creators Cathy O'Neil
146. Thanks!