Upgrade to Pro — share decks privately, control downloads, hide ads and more …

You Can Do Deep Learning!

You Can Do Deep Learning!

It’s 2018, and you don’t need a PhD to do deep learning. This talk will empower you to build your first deep learning project using Python and open-source tools. In other domains, like web development, we leverage frameworks that give us high-level abstractions to work with, and I will show you that deep learning doesn’t have to be different.

A1225054c221cedf3635ba79582c8768?s=128

William Horton

July 28, 2018
Tweet

More Decks by William Horton

Other Decks in Technology

Transcript

  1. You Can Do Deep Learning! By William Horton

  2. None
  3. Image Classifier cat? dog?

  4. None
  5. Other applications

  6. Who am I? • Backend Engineer, Data Team at Compass

    • Python • Deep Learning
  7. None
  8. Let’s dive in!

  9. None
  10. Deep Learning

  11. None
  12. None
  13. None
  14. $ gem install rails $ rails new blog $ cd

    blog $ bin/rails server $ pip install Django $ django-admin startproject blog $ cd blog $ python manage.py runserver
  15. None
  16. None
  17. Magic!

  18. Deep Learning

  19. None
  20. Common roadblocks What about the math? Having to learn everything

    about everything before you even let yourself start Thinking you can’t do this because you don’t have a PhD
  21. None
  22. the fast.ai formula:

  23. Math is important Theory is important But why not start

    with the knowledge and skills that you have?
  24. https://cloud.google.com/blog/big-data/2016/08/how-a-japanese-cucumber-farmer-is- using-deep-learning-and-tensorflow

  25. Use what you know!

  26. None
  27. None
  28. You know how to translate ideas into working code.

  29. None
  30. You know how to read the documentation and source of

    libraries.
  31. None
  32. You’ve probably spent hours of your life debugging.

  33. None
  34. Putting it together

  35. None
  36. Kaggle Kernels

  37. None
  38. fast.ai library

  39. Communities fast.ai forums Pytorch forums Kaggle discussion threads AI Saturdays

    Medium Twitter
  40. http://tools.google.com/seedbank/

  41. What can I do?

  42. Near term questions What happens if I train for longer?

    What happens if I change the learning rate? Can I apply this to another dataset? Can I use a different architecture? What is training loss? What is validation loss?
  43. Understand, implement, and explain https://medium.com/@radekosmulski/do-smoother-ar eas-of-the-error-surface-lead-to-better-generalization- b5f93b9edf5b http://teleported.in/posts/cyclic-learning-rate/

  44. https://medium.com/@hortonhearsafoo/adding-a-cutting-edge-deep-learning-train ing-technique-to-the-fast-ai-library-2cd1dba90a49

  45. Win an Emmy?

  46. None
  47. None
  48. Tackle your business problems

  49. Take on the giants https://www.theverge.com/2018/5/7/17316010/fast-ai-speed-test-stanford-dawnbe nch-google-intel

  50. None
  51. None
  52. Now get started! (And let me know: @hortonhearsafoo)

  53. Image Credits Terminator: https://www.sideshowtoy.com/assets/products/300157-terminator-t-800-endoskeleton/lg/terminator-2-terminator-t-800-endoskeleton-m aquette-sideshow-300157-17.jpg Dog: https://i.ytimg.com/vi/SfLV8hD7zX4/maxresdefault.jpg Cat: https://upload.wikimedia.org/wikipedia/commons/thumb/3/3a/Cat03.jpg/1200px-Cat03.jpg ImageNet

    graph: https://www.economist.com/special-report/2016/06/25/from-not-working-to-neural-networking Google Translate: https://translate.google.com/ AlphaGo: https://deepmind.com/research/alphago/ Compass: https://www.compass.com/ Tensorflow: screenshotted from https://www.tensorflow.org/ Coursera: from https://www.facebook.com/Coursera/ Khan Academy: https://cdn.kastatic.org/images/khan-logo-dark-background.png Kaggle: https://en.wikipedia.org/wiki/Kaggle Udacity: https://d20vrrgs8k4bvw.cloudfront.net/images/open-graph/udacity.png Roadblock: https://thesandwichedman.files.wordpress.com/2017/03/roadblock.jpg?w=519 You Shall Not Pass: https://coldcallcoach.net/donts-grappling-gatekeeper/ django: https://en.wikipedia.org/wiki/Django_(web_framework) Rails: screenshotted from https://www.youtube.com/watch?v=Gzj723LkRJY DHH: http://david.heinemeierhansson.com/images/me.jpg
  54. You’re on rails: screenshotted from https://guides.rubyonrails.org/getting_started.html Magic: https://recruitingtools.com/wp-content/uploads/sites/2/2017/07/magic-wand.jpg fastai: http://www.fast.ai/

    fastai video: https://www.youtube.com/watch?v=Th_ckFbc6bI fastai formula: screenshotted from http://course.fast.ai/ python: https://www.python.org/ jira ticket: http://www.taigeair.com/JIRA-Ticket mockup: https://wireframesketcher.com/sample-mockups.html keras documentation: https://keras.io/getting-started/functional-api-guide/ pytorch github: https://github.com/pytorch/pytorch stack trace: http://i45.tinypic.com/nzm99h.jpg stackoverflow: https://stackoverflow.com/ gpu: https://www.notebookcheck.net/fileadmin/_processed_/d/3/csm_GeForce_GTX_1080ti_3qtr_top_left__f25b948c6c.jpg data: http://metaltechalley.com/wp-content/uploads/2017/09/data.jpg jupyter: http://jupyter.org/assets/try/jupyter.png tensorflow: https://upload.wikimedia.org/wikipedia/commons/thumb/1/11/TensorFlowLogo.svg/2000px-TensorFlowLogo.svg.png keras: https://keras.io/ scikit-learn: https://twitter.com/scikit_learn pytorch: https://pytorch.org/ emmy nominations: http://www.emmys.com/awards/nominees-winners/2018/outstanding-creative-achievement-in-interactive-media-within-a-scripted-progr am not hotdog app: http://www.bgr.in/news/the-ridiculous-not-hotdog-app-is-real-and-here-are-6-other-apps-if-you-love-it/ not hotdog linkedin post: https://www.linkedin.com/feed/update/urn:li:activity:6423300978275082240/ dawnbench: https://dawn.cs.stanford.edu/benchmark/