Presented by Annie Flippo, Manager of Analytics and Lead Data Scientist, Thinknear by Telenav

1. Hot Dog Not Hot Dog!

2. Generate new training data
Without taking more photos
Annie Flippo
October, 2017
Hot Dog, Not Hot Dog!

3. Image Source: HBO

4. Image Source: HBO

5. Image Source: HBO

6. The Tech
is good
Just needs
more training
data
Image Source: HBO

7. Image Source: HBO

8. I work as a Manager of Analytics
At Thinknear by Telenav

9. We are a location-based advertising platform

10. GPS navigation for vehicles and mobile phones

11. Cat
Dog
Mug
Hat

12. Deep Neural Network

13. Neural Network Visualized
http://scs.ryerson.ca/~aharley/vis/conv/

14. Perfect Training Data

15. Real-life Data

16. Hot Dogs or Meatballs?
Image recognition is one of the most important
developments in AI. Close your eyes and imagine:
- Taking a shower
- Eating your lunch
- Driving

17. Real-life Driving: Distracted

18. Our Eyes Only

19. Autonomous Driving: Sign Recognition

20. Real Signs

21. Why is object recognition so hard?
Image Source: Stanford University, cs231n.github.io/classification

22. The Humble Stop Sign

23. Be Aware of Image Hacking 1
Image Source: https://codewords.recurse.com/issues/five/why-do-neural-networks-think-a-panda-is-a-vulture

24. Be Aware of Image Hacking 2
Image Source: https://codewords.recurse.com/issues/five/why-do-neural-networks-think-a-panda-is-a-vulture

25. Be Aware of Image Hacking 3
Image Source: https://medium.com/@ageitgey/machine-learning-is-fun-part-8-how-to-intentionally-trick-neural-networks-b55da32b7196

26. Be Aware of Image Hacking - Fake Fakes
This could be a devastating example if the autonomous vehicle is trained
to drive over newspaper!
Pedestrian (98.5%) Newspaper (95.9%)

27. What can we do?
I need better training!
Any ideas?

28. Generate Training Data Idea 1
Augment source image by:
Rotation Perspective Transformation

29. Generate Training Data Idea 2
Blurring Brightness Adjustment

30. Generate Training Data Idea 3
Add NoiseShearing

31. Apply Blurring

32. Apply Noise

33. Adjust Brightness

34. Apply Rotations

35. All Adjustments

36. Dropout in Neural Network

37. Time Saving Tip
Generate many variations per source image
but:
● Don’t save to disk (slow and takes space)
● Run it under a Python Generator Object
○ Generate images as you train (fast and scalable)

38. Train a Winning Model
… and become a billionaire!
Image Source: HBO

39. Thank You
acflippo @acflippo
Slides and code: bit.ly/scds-acflippo