Deep Learning Talk - Saverin

Deep Learning Introduction Talk @ Saverin

Transcript

1. Deep Learning Yasser Souri - Alireza Nourian http://sobhe.ir
2. Have you ever heard of ... Neural Networks
3. Have you ever heard of ... Deep Learning
4. Who is he?
5. Who is he? Jeff Dean, Google
6. Jeff Dean Creator of Map Reduce, Big Table, Google Crawler
7. Jeff Dean Creator of Map Reduce, Big Table, Google Crawler Google Ads, Google Translator, ...
8. Jeff Dean Facts Compilers don't warn Jeff Dean. Jeff Dean warns compilers.
9. Jeff Dean’s Calculator
10. Jeff Dean’s Current Role Google Brain
11. DeepMind In 2014, Google acquired DeepMind (a team of ~50) for ~$ 500 million. And facebook wanted to buy them also.
12. What is Machine Learning? Problem 1: Given a sequence of numbers, sort them
13. What is Machine Learning? Problem 1: Given a sequence of Farsi characters, output Pinglish
14. What is Machine Learning? Problem 3: Give a grayscale 28x28 pixel image, identify what number it is.
15. What is Machine Learning? Problem 3: Give a grayscale 28x28 pixel image, identify what number it is.
16. What is Machine Learning? x f(x) y Classic
17. What is Machine Learning? x f(x) y g(x) y’ h(x) y” Classic
18. How to Solve Machine Learning Problems Data = (x, y) Classic
19. How to Solve Machine Learning Problems Data = (x, y) y = f(x) Classic (x, y) f(x)
20. How to Solve Machine Learning Problems Data = (x, y) y = f(x) Learn the parameters Classic (x, y) f(x; w)
21. How to Solve Machine Learning Problems Data = (x, y) y = f(x) Learn the parameters Can x be the raw pixels? Classic (x, y) f(x; w) Features
22. How to Solve Machine Learning Problems Data = (x, y) y = f(x) Learn the parameters Can x be the raw pixels? Classic (x, y) f(x; w) Features O(#features) ~ O(#parameters)
23. Machine Learning Demo http://playground.tensorflow.org/ Classic
24. Deep Learning Basics Learn from raw data y = f(g(h( … (x) ))) Deep
25. Deep Learning Learn from raw data Number of parameters are much larger y = f(g(h( … (x) ))) Deep
26. Deep Learning Learn from raw data Number of parameters are much larger You need more data to learn y = f(g(h( … (x) ))) Deep
27. Problems being solved with deep learning Deep
28. Problems being solved with deep learning Deep
29. One to one: Image Classification Deep
30. One to one: Image Classification Deep
31. Problems being solved with deep learning
32. One to Many: Image Captioning Describing Images:
33. Fun With ConvNets Describing Images:
34. Problems being solved with deep learning
35. May to One: Generating Images Generating Images:
36. May to One: Generating Images Generating Images:
37. Problems being solved with deep learning
38. Statistical Machine Translation
39. End-to-End Neural Machine Translation (1) Hirschberg, J. & Manning, C. D. Advances in natural language processing, Science, 2015, 349, 261-266
41. Learning to Execute
42. Deep Reinforcement Learning
43. Demo Videos https://www.youtube.com/watch?v=ePv0Fs9cGgU https://www.youtube.com/watch?v=Q70ulPJW3Gk
44. Fun With ConvNets Modifying images:
45. Fun With ConvNets Style transfer:
46. Fun With ConvNets Style transfer:
47. Fun With ConvNets Colorization:
48. Fun With ConvNets Colorization:
49. Fun With ConvNets Colorization:
50. Fun With ConvNets Colorization:
51. Fun With ConvNets Colorization:
52. Fun With ConvNets Colorization:
53. Fun With ConvNets Colorization:
54. Growing Use of Deep Learning at Google Jeff Dean & Oriol Vinyals, “ Large Scale Distributed Systems for Training Neural Networ”, NIPS 2015.
55. Deep Learning Tools