Machine learning: How it can help your business - Microsoft Future Decoded - Budapest, March 2018

Transcript

1. Machine Learning: How It Can Help Your Business Szilárd Pafka,

   PhD Chief Scientist, Epoch (USA) Microsoft Future Decoded, Budapest March 2018
2. None

3. Disclaimer: I am not representing my employer (Epoch) in this

   talk I cannot confirm nor deny if Epoch is using any of the methods, tools, results etc. mentioned in this talk
4. None

5. Source: Andrew Ng

6. None

7. y = f(x) “Learn” f from data Source: Hastie etal,

   ESL 2ed

8. Machine Learning linear/logistic regression decision trees neural networks support vector

   machines random forests gradient boosting deep learning neural networks

9. Machine Learning linear/logistic regression (early 1900s/60s) decision trees (60s/80s) neural

   networks (60s/80s) support vector machines (90s) random forests (90s) gradient boosting (90s) deep learning neural networks (2000s)
10. None

11. data mining Source: Szilard Pafka

12. data science Source: Szilard Pafka

13. data science Source: Szilard Pafka

14. CRISP-DM, 1999

15. data $$$

16. How?

17. None
18. None

19. Source: Andrew Ng

20. None
21. None
22. None

23. Source: @iamdevloper (twitter)

24. None
25. None
26. None

27. structured/tabular data: GBM (or RF) very small data: LR very

    large sparse data: LR with SGD images/videos, speech: DL

28. structured/tabular data: GBM (or RF) very small data: LR very

    large sparse data: LR with SGD images/videos, speech: DL better answer: it depends

29. structured/tabular data: GBM (or RF) very small data: LR very

    large sparse data: LR with SGD images/videos, speech: DL better answer: it depends alternative answer: try them all

30. structured/tabular data: GBM (or RF) very small data: LR very

    large sparse data: LR with SGD images/videos, speech: DL better answer: it depends alternative answer: try them all extra accuracy: combine them (ensembles)
31. None
32. None

33. 10x

34. None
35. None
36. None

37. ML training: lots of CPU cores lots of RAM limited

    time
38. None
39. None
40. None
41. None
42. None

43. Source: Szilard Pafka

44. None

45. Random forest GBM GBM + cross validation GBM + hyperparameter

    tuning Logistic regression Neural Nets / Deep Learning Ensembles
46. None
47. None
48. None
49. None
50. None
51. None

52. Backup Slides

53. None

54. 10x

55. None
56. None
57. None

58. Source: Szilard Pafka: 10 Pitfalls in Data Science, LA Data

    Science Meetup, February, 2014
59. None
60. None