Theano: let's build a net

Transcript

1. Theano: let’s build a net Ekaterina Tuzova

2. Motivation

3. MNIST dataset X — images (60000, 784) Y — labels

   (60000, ) Goal: train a model to look at images and predict what digits they are.

4. K nearest neighbors

5. K nearest neighbors ~95 % accuracy

6. Feature engineering

7. Machine Learning

8. Machine Learning

9. Variable in Computer Science A piece of informations stored somewhere

   in memory with a name made up by a programmer.

10. Variable in Mathematics Indeterminate quantity that can be manipulated using

    arithmetic operations, or using calculus and so on.

11. Theano

12. What is Theano? A library for efficiently handling mathematical expressions

    involving tensors.

13. What is tensor? Multi-dimensional array.

14. What is Theano? A mathematical symbolic expression compiler.

15. Quick How-To - Define symbolic expression - Compile a function

    that can compute numeric values using those expressions - Execute that function on data

16. Example model GoogLeNet, 2014 ~99.6% accuracy

17. Logistic regression Probabilistic, linear classifier.

18. Simple model

19. Preparation

20. Load MNIST

21. Define model

22. Theano shared variable Shared between different functions and function calls.

23. Build model

24. Train model

25. Train model ~93% accuracy

26. Simple model

27. Improve model

28. Improve model

29. Improve model

30. Train model ~98.5% accuracy

31. Thank you. @ktisha