Uploaded byindico data
The Unreasonable Benefits of Deep Learning
The document discusses deep learning, highlighting its evolution and significance in various applications such as image recognition, natural language processing, and sentiment analysis. It emphasizes the advantages of deep learning, including simplicity, accuracy, and flexibility, compared to traditional machine learning and expert systems. The text also notes that deep learning can learn from vast amounts of data and makes abstraction possible, posing the question of what problems deep neural networks can solve.
More Related Content
Similar to The Unreasonable Benefits of Deep Learning
![Getting started with indico APIs [Python]](https://cdn.slidesharecdn.com/ss_thumbnails/papisconference-141204130351-conversion-gate01-thumbnail.jpg?width=640&height=640&fit=bounds)
![[Redis Released]- FalkorDB - Redis + Graph Agentic Memory’s Secret Sauce](https://cdn.slidesharecdn.com/ss_thumbnails/redisreleased-falkordbslidedeck-1125-251115194922-e1c0046b-thumbnail.jpg?width=640&height=640&fit=bounds)
The Unreasonable Benefits of Deep Learning
- 1. The Unreasonable Benefits ofDeep Learning Daniel Kuster, Ph.D. @djkust @indicodata
- 2. “All good researcherswill tell you that the most promising direction is the one they are currently pursuing. If they thought something else was more promising, they would be doing that instead.” — G. Hinton
- 3. What is deeplearning? …a method for applying simple mathematical functions to data.
- 4. web: search, facialrecognition smartphones: speech -> text email: smart reply mail: handwriting -> digits cars: pedestrian detection art & design: artistic style transfer
- 5. Wait, why now? ~1960’s(visual cortex is a deep neural network) Simple neurons ⇾ hierarchical features ⇾ complex ~1990’s (computational models) Neural networks ⇾ simple functions, applied piecewise ~ 2000’s (the internet + cheap storage) Lots of data ~2012 (2 GPUs beat Google’s 16,000 CPU cluster) Very fast and cheap parallel computing power Deep neural networks ≈ mathematics + data
- 6. How deep learningworks: (1-minute theoretical explanation)
- 7. Cat detector =eyes + fur + nose + … …but how do we discover the features? and where are they?
- 8. input simple math functions pooling simplemath functions classifier But how do we know what to detect? learn from the data…
- 9. How to usedeep learning: (10-second practical explanation)
- 10. data (content) model classifier data (predictions) Howto use deep learning
- 11. Maybe you areskeptical that deep learning will have a lasting impact…
- 12. Maybe you areskeptical that mathematics will have a lasting impact?
- 13. “The enormous usefulnessof mathematics in the natural sciences is something bordering the mysterious and there is no rational explanation for it.” —Eugene Wigner (1960) “The Unreasonable Effectiveness of Mathematics in the Natural Sciences” Unreasonable Benefits of Deep Learning
- 14. “…mathematical formulation…leads inan uncanny number of cases to an amazingly accurate description of a large class of phenomena.” “…the concepts of mathematics are not chosen for their conceptual simplicity…but for their amenability to clever manipulations and to striking, brilliant arguments.” —Eugene Wigner (1960) “The Unreasonable Effectiveness of Mathematics in the Natural Sciences” Unreasonable Benefits of Deep Learning
- 15. simplicity Unreasonable Benefits ofDeep Learning accuracy flexibility hacks
- 16. simplicity Unreasonable Benefits ofDeep Learning accuracy flexibility hacks
- 17. Simple? Compared towhat? • Expert systems Domain expertise ⇾ think a lot ⇾ codify rules (e.g., 1700 pages of English grammar) More data, more pain. Previous wave of “A.I.” (good rules can seem magical). • Traditional machine learning Data ⇾ domain expertise ⇾ feature extraction ⇾ learned weights Learn everything from scratch. Manual feature engineering, biased and tedious. More data helps! • Deep neural networks Data ⇾ model ⇾ learned weights End-to-end learning, directly from examples. Like we (humans) do. Can learn transferable features. More data really helps!
- 18. Example: the simplesttext analysis task: sentiment!
- 19. text (reviews) model classifier Sentiment: compress textto one bit of info 1 0
- 20. text (reviews) model classifier Sentiment: compress textto one bit of info
- 21. Sentiment over time: theshapes of stories
- 22. story + vis+ code http://indico.io/blog/plotlines
- 23. text model classifier Emotion: compresstext to __ bits of info 😀 😉 😡 😈 😭 😍 😎 😞
- 24. Emotion import indicoio indicoio.emotion(“What?!?!? Ihad no idea this sort of thing existed!”) (~two lines of code)
- 25. text model classifier Personality,topics, political lean, language, … compress text to __ bits of info
- 26. image model classifier imagefiltering: compress images to one bit of info mature content safe (for work)
- 27. Content filtering (especially foruser-generated content) You have a brand Your brand has an identity (Disney vs. Calvin Klein) Your audience might have different sensibilities than you do, about what is appropriate for your brand Filter out the inappropriate content at your own custom threshold
- 28. simplicity Unreasonable Benefits ofDeep Learning accuracy flexibility hacks
- 29. Most accurate sentimentAPI: (93.8% on IMDB)
- 30. OK, accuracy isgreat… …but how does this help me solve problems faster/cheaper/better?
- 31. simplicity Unreasonable Benefits ofDeep Learning accuracy flexibility hacks
- 32. Access features directly …feedinto a new classifier data (content) model classifier data (predictions)
- 38. Get the code(free): https://github.com/IndicoDataSolutions/SuperCell
- 39. Labeling 100k+ examples…sucks! …labelinga few hundred is just a couple hours at the coffeeshop.
- 40. data (content) model classifier data (predictions) What’shappening in the the middle? Let’s look at some features
- 41. RNN for sentimentprediction t-SNE of recurrent features: pos/neg words
- 42. Features learned somerules of English from binary sentiment labels
- 43. “The big payoffof deep learning is to allow learning higher levels of abstraction” — Y. Bengio
- 44. simplicity Unreasonable Benefits ofDeep Learning accuracy flexibility hacks!
- 45. Features are compressedknowledge Who says we can’t combine them?
- 46. Experiment: image features +text features A man standing in a field holding a small parachute image encoder text encoder similarity(image, text)
- 47. “in the sky”⇾ most similar images
- 48. photo with wineglass ⇾ intent ⇽ campaign image-in-image search
- 49. Q: What problemscan be solved with a deep neural network? A: If a human mind can do it in 1/10th of a second, a deep neural network can probably do it well enough… assuming you have data!
- 50. “Many scientists (myselfincluded) take a sadistic pleasure in proving other people wrong. — Y. LeCun
- 51. The Unreasonable Benefitsof Deep Learning: simplicity, accuracy, flexibility, hacks Questions? Daniel Kuster, Ph.D. @djkust @indicodata
- 52. Image credits: Unsplash (backgrounds) Googlesearch, Facebook AI Research (DeepFace), nVidia, Gatys et al. (arXiv: 1508.06576) Brigitewear International (Borat swimsuit) Imgur (skeptical dogs) Jack the cat …and the team at indico!
- 53. “A machine learningresearcher, a crypto-currency expert, and an Erlang programmer walk into a bar. Facebook buys the bar for $27 billion.” @ML_Hipster