Introduction to Machine Learning

Transcript

1. introduction to machine learning Pierre Margueritte (@mfpierre) lead engineer at
2. let’s solve a problem vs
3. function isOrangeOrApple(image) { // detect color ? // detect edges ? // detect textures ? // analyze shape ? // compute probabilities } traditional approach
4. did you think of all cases?
5. hard coded rules do not scale well
6. learning from examples and experience
7. examples = data sets )PSTFQPXFS 4FBUT TQPSUTDBS   TQPSUTDBS   DBS   DBS   features classes
8. choosing an algorithm
9. building a classifier from sklearn import tree features = [[300, 2], [350, 2], [120, 5], [150, 4]] labels = ["sports-car", "sports-car", "car", "car"] clf = tree.DecisionTreeClassifier() clf = clf.fit(features, labels) print clf.predict([[200, 4]]) [‘car’]
10. http://i.imgur.com/gBKH3cj.gif
11. under the hood horsepower <= 225.0 gini = 0.5 samples = 4 value = [2, 2] class = car gini = 0.0 samples = 2 value = [2, 0] class = car True gini = 0.0 samples = 2 value = [0, 2] class = sports-car False
12. the data you learn from is key
13. analyzing a feature IPSTFQPXFS     DBS TQPSUTDBS probably a car probably a sports-car
14. so much features
15. features you want to avoid DPMPS
    PG
    UIF
    DBS SFE ZFMMPX HSFZ CMBDL DBS TQPSUTDBS probably…
16. feature engineering
17. feature engineering
18. what makes a good feature? informative independent simple to understand
19. but how do I know If my predictions are actually good?
20. splitting your data set training test 80 % 20 %
21. the confusion matrix QSFEJDUJPO
    /0 QSFEJDUJPO
    :&4 BDUVBM
    /0 5/ '1 BDUVBM
    :&4 '/ 51
22. machine learning is a lot about experimentation
23. a word on deep learning
24. artificial neural networks
25. why deep?
26. scale made neural networks exciting
27. how we use ML at
28. 75% of our text content is moderated automatically (that’s 7M items/month)
29. text moderation
30. we’re only at the beginning!
31. questions?