Teaching Machine Learning

Insights on how to teach machine learning and deep learning.
The entry barrier is not that high!
Video: https://www.youtube.com/watch?v=dyoxtDhUR74
Conference: https://pydata.org/warsaw2017/

Transcript

1. Teaching
   Machine
   Learning
   Piotr Migdał, PhD
   http://p.migdal.pl @pmigdal
   PyData Warsaw Conference
   19 Oct 2017
   

   View Slide

2. PhD in quantum physics theory
   (2014, ICFO, Barcelona)
   data scientist
   (deepsense.ai / consultant)

   machine learning
   deep learning
   data-viz (D3.js)
   

   View Slide

3. I teach…
   

   View Slide

4. View Slide

5. Outline
   • Everything is easy
   • Typical problems
   • Getting pragmatic
   

   View Slide

6. View Slide

7. A person first learns
   classical mechanics
   • …by playing with balls, blocks?
   • …by learning Newton laws, differential calculus?
   

   View Slide

8. A person first learns
   natural numbers
   • …by counting apples, toys?
   • …by the von Neumann construction?
   

   View Slide

9. i
   ~ ˙ =
   ⇣
   ~2
   2m
   r2 +
   V
   (
   x
   )
   ⌘
   ˆ
   H = E
   

   View Slide

10. A person first learns
    quantum mechanics
    • ...by learning linear algebra, complex numbers?
    

    View Slide

11. http://p.migdal.pl/2016/08/15/quantum-mechanics-for-high-school-students.html
    

    View Slide

12. http://quantumgame.io/
    

    View Slide

13. A person first learns
    machine learning
    • ...by studying computer science, mathematics

    and statistics for years?
    

    View Slide

14. you don’t understand
    Machine Learning
    unless you can teach it with pen&paper
    (or at least - JavaScript)
    

    View Slide

15. View Slide

16. https://generalabstractnonsense.com/2017/03/A-quick-look-at-Support-Vector-Machines/
    

    View Slide

17. Machine Learning
    https://twitter.com/b0rk/status/821922905890103298/photo/1
    

    View Slide

18. Decision trees, visually
    http://www.r2d3.us/visual-intro-to-machine-learning-part-1/
    

    View Slide

19. http://p.migdal.pl/2017/01/06/king-man-woman-queen-why.html
    Julia Bazińska’s talk “Exploring word2vec vector space" - tomorrow at 15:00
    Word2vec
    

    View Slide

20. Matrix factorizaton
    http://p.migdal.pl/matrix-decomposition-viz/, work in progress
    

    View Slide

21. https://distill.pub/
    

    View Slide

22. …but at least
    Deep Learning
    is hard, isn’t it?
    

    View Slide

23. View Slide

24. View Slide

25. View Slide

26. Spreadsheet-based
    deep learning
    http://www.deepexcel.net/
    

    View Slide

27. http://setosa.io/ev/image-kernels/
    

    View Slide

28. Things which are not
    problems
    • I don’t know Python
    • They are only high-school
    • They are not from STEM
    

    View Slide

29. Trypophobia detector
    by high-school students
    https://github.com/cytadela8/trypophobia
    

    View Slide

30. Trypophobia detector
    by high-school students
    https://github.com/cytadela8/trypophobia
    

    View Slide

31. Problems with teachers
    • Too much math details & too little insight
    • Too much historical inertia
    • No plots
    • Too little real data

    (e.g. all np.random.randn(n, m))
    

    View Slide

32. Real data, plots > arrays
    

    View Slide

33. Pragmatic algorithms
    • kNN
    • Linear + Logistic

    Regression
    • Random Forest
    • XGBoost
    • Neural Networks
    

    View Slide

34. Problems with clients
    • Everyone is an “expert”
    • Squeezing one semester (or a few)

    into a few days (or just one)
    • Deep learning will solve all our problems
    • Installation!!!
    

    View Slide

35. Logistic Regression

    vs Random Forest
    vs Deep Learning
    https://github.com/szilard/benchm-ml
    http://datascience.la/benchmarking-random-forest-implementations/
    

    View Slide

36. Easy setup
    • Python 3 with Anaconda
    • Jupyter Notebook
    • Neptune.ML
    

    View Slide

37. Thank you!
    Questions?
    more on my blog

    http://p.migdal.pl/ @pmigdal
    

    View Slide