GAN for business value @ Data Science Milan
Download as PPTX, PDF
1 like287 views
This document discusses generative modeling beyond using GANs to generate creepy images on Twitter. It provides examples of using generative modeling for tasks like generating video game levels, text, molecules, and music. Generative modeling involves modeling a distribution to generate new data points, while discriminative modeling involves classifying data points. Generative modeling has value in data augmentation, anonymization, and translating data from one domain to another.
GAN for business value @ Data Science Milan
- 1. Generative modeling beyond creepy GIFs in Twitter Alexandr Honchar
- 2. AI architect @ Mawi Solutions AI consultant @ CIS, EU Masters in maths candidate @ University of Verona Blogger @ Medium
- 7. Random artists from Twitter, Telegram
- 11. But I dont need all that! I am serious guy who needs data science to earn money, not Twitter likes
- 12. Generative vs discriminative ML A bit of scary mathematics (no)
- 13. Space of cats and dogs A very deep neural network already created some nice embedding space for us from initial pixel space
- 14. Discriminative modeling: A function f(x, w) telling us where new x belongs to from {cat, dog}
- 15. Generative modeling: A modeled distribution P(x|y), where x - our data point and y belongs to {cat, dog} So, we can or create new cat from P(x|y=cat), or check if some x_i belongs to P(x_i|y=dog) using well known maths
- 16. Generative modeling: A modeled distribution P(x|y), where x - our data point and y belongs to {cat, dog} So, we can or create new cat from P(x|y=cat), or check if some x_i belongs to P(x_i|y=dog) using well known maths
- 17. Natural manifold hypothesis: real-world high dimensional data (such as images) lie on low- dimensional manifolds embedded in the high-dimensional space Short tails, a little fur Long tails, a lot of fur
- 18. Not only images!
- 19. Video game level generation Text generation Drug molecule generation, Insilico Music generation, Google Magenta
- 20. P(X) What value has generation of objects?
- 22. The RAPIDD Ebola forecasting challenge: Model description and synthetic data generation GAN-based Synthetic Medical Image Augmentation for increased CNN Performance in Liver Lesion Classification IoT data generation, aidrome
- 23. Short tails, a little fur Long tails, a lot of fur
- 24. TL-GAN: transparent latent-space GAN
- 25. Medical Image Synthesis for Data Augmentation and Anonymization using Generative Adversarial Networks Numerai / Erasure for financial data prediction
- 26. ~P(X) What are all other objects?
- 29. P(X) -> P(Y) What value has turning some objects to another ones?
- 31. CycleGAN
- 34. Short tails, a little fur Long tails, a lot of fur
- 35. Voice from the phone <-> Voice from HQ mic, SEGAN ECG from hospital <-> ECG from a wristband, Mawi Official text <-> Funny text, MaskGAN Dressed person <-> Naked person, DeepFakes CCTV camera <-> HD camera
- 36. 8x8 to 128x128 Super Resolution with Adversarial Autoencoders
- 37. Takeaways
- 38. Generative modeling >= Discriminative modeling I do not understand things which I cannot create We need to share data to third parties We need to manipulate our data with simple factors We need anomaly detectors and reject options We need to adapt our data for some conditions
- 39. FACEBOOK: @rachnogstyle MEDIUM / TWITTER: @alexrachnog alex@mawi.band