Machine Learning APIs

Transcript

1. Machine Learning APIs Christian Thurau christian.thurau@unbelievable-machine.com
2. What is Machine Learning?
3. Classification: Face Detection
4. 23.04.15 The unbelievable Machine Company 4 Use data to …
5. …train a model…
6. … apply model to unknown data
7. Some more recent examples
8. Skype: Real-timeTranslation
9. Google: Natural Description of Images
13. Machine Learning APIs
15. Pre-trained models & APIs Text: •  Sentiment analysis (AlchemyAPI, Indico, DatumBox) •  Spam detection (DatumBox) •  Language detection (DatumBox) •  Topic classification (DatumBox) Images: •  Image tagging (clarifai, AlchemyAPI, imagga) •  Face detection (AlchemyAPI) •  Facial emotion recognition (Indico)
16. Pre-trained models & APIs Text: •  Sentiment analysis (AlchemyAPI, Indico, DatumBox) •  Spam detection (DatumBox) •  Language detection (DatumBox) •  Topic classification (DatumBox) Images: •  Image tagging (clarifai, AlchemyAPI, imagga) •  Face detection (AlchemyAPI) •  Facial emotion recognition (Indico) No training data or ML knowledge required J
17. Flexible ML APIs •  Google Prediction API •  Microsoft Azure •  Amazon Machine Learning API •  IBM Watson •  BigML •  Prediction.io (open source J)
18. Flexible ML APIs •  Google Prediction API •  Microsoft Azure •  Amazon Machine Learning API •  IBM Watson •  BigML •  Prediction.io (open source J) Little ML knowledge required J You can use your own data J
19. However, training a model is seldom sufficient …
20. …It’s the features, stupid!
21. Training Data Example from „Image Processing and Computer Vision“, Golan Levin E.g. for images:
22. Training Data Features Example from „Image Processing and Computer Vision“, Golan Levin E.g. for images:
23. Training Data Features Model training Example from „Image Processing and Computer Vision“, Golan Levin Model training E.g. for images:
26. Good feature!
28. Decision Boundary
29. Bad feature! Decision Boundary
30. Pretty bad image feature Feature.. •  extraction •  engineering •  selection •  preprocessing
31. Pretty bad image feature Pretty good image feature Feature.. •  extraction •  engineering •  selection •  preprocessing The success of ML depends on your data representation
32. How to find good features? •  Use domain knowledge •  For some problems: Deep Learning to the rescue! •  For all other problems: –  ask a ML expert –  scientific papers But: this is (usually) not part of ML APIs L
33. Conclusion •  Machine learning APIs are great if: –  the problem is generic (machine translation) –  the problem has been solved (face detection) •  Machine learning APIs are difficult to use if: –  you have to use your own data (feature engineering) –  the problem is new or unfeasible (stock market prediction) •  Customized ML as a service solutions : –  preprocess and clean data –  feature engineering –  find & train the right model
34. *um E christian.thurau@unbelievable-machine.com T +49 - 30 - 8892656 - 694 M +49 - 152 - 22 8 44 205 Berlin, Germany The unbelievable Machine Company GmbH