際際滷

際際滷Share a Scribd company logo

Deep learning: from theory via virality to applications in business

Download as PPTX, PDF
R
Rasmus Rothe

This talk was held by Rasmus Rothe at the Berlin Machine Learning Group at Rocket Internet on June 2, 2016. He presented the research behind howhot.io, some fun facts, how it became viral and how he intends to leverage the technology behind it to solve real-world problems with his new company, Merantix.

1 of 26
Downloaded 31 times
Deep learning: from theory via virality to applications in business Berlin Machine Learning Group Rocket Internet Rasmus Rothe June 2, 2016
Deep learning: from theory via virality to applications in business
16-18 SEPTEMBER 2016
Howhot.io
Deep learning: from theory via virality to applications in business
Deep learning: from theory via virality to applications in business
Deep learning: from theory via virality to applications in business
Deep learning: from theory via virality to applications in business
100m+ evaluations to date 1m visitors in first 12hrs after launch Spikes in usage due to media attention
Its more than a joke Rasmus Rothe, Radu Timofte, and Luc Van Gool. "DEX: Deep EXpectation of apparent age from a single image., ICCVW. 2015. Winner of LAP challenge on apparent age estimation (115 teams) NVIDIA ChaLearn LAP 2015 Best Paper Award Rasmus Rothe, Radu Timofte, and Luc Van Gool. "Some like it hot- visual guidance for preference prediction., CVPR, 2016 Rasmus Rothe, Radu Timofte, and Luc Van Gool. Deep expectation of real and apparent age from a single image without facial landmarks, in submission IMDB-WIKI dataset publicly available: https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/
Rasmus Rothe Radu Timofte Luc Van Gool Alex Zimmermann Jan Berchtold How it started
120,000 profile images 22 million hot-or-not ratings 44.81% mens ratings positive 8.58% womens ratings positive Dataset statistics
Datasource # Photos Gender (m/f) Age (23.4yrs) Attractiveness (top 23%)IMDb 500,000 x x Wikipedia 60,000 x x Blinq 120,000 x x x 1) Data to learn from
2) Teaching the deep neural network
3) What happens when you upload a photo
There is more research to it ?
Prediction RatersVisual space Nearest neighbor Collaborative filtering PT Q 4.0 8.4 2.5 9.7 9.1 3.2 7.7 7.6 6.0 1.5 0.5 8.7 6.0 0.8 2.5 8.5 0.8 8.2 8.9 3.8 8.1 8.2 1.8 7.0 6.7 3.9 7.4 0.8 0.8 5.4 3.9 3.8 8.2 1.8 7.0 6.7 3.9 7.4 5.9 4.0 Visual QueryPastratings OR Matrix factorization Latent space (Q) Regression
What led to winning the age estimation challenge (115 teams)? 1. Deeper models (VGG-16) 2. Ensemble classification (20 models) 3. More training data (500k+ images) 5. Robust face alignment 6. Deep expectation layer 4. Augmentation
Deep learning: from theory via virality to applications in business
Global Online Players Which of these companies was founded in Germany? Exception:
If we look closer 1600out of 1900 of Germanys biggest companies, are B2B Industry Companies. Source: Venohr, B., Fear, J., & Witt, A. (2015). "Best of German Mittelstand- The world market leaders.", Statista V.S. Thats what we are excellent at and lightyears ahead of Silicon Valley. x8 turnover!
Merantix is an AI Research Laband a Company Builderthat helps industry companies to leverage the power of AI. Rethinking the German core strengths ENGINEERING
Why focus on B2B industries? Industry 1 2 27 201 230 384 B2C Chatbots Gaming Medical Technology Chemicals Machinery & Equipment Automotive Consumer Turner over in EUR in Germany - Sources: Germany Trade and Invest
Why applying generic toolboxes is not enough 1. Dirty data 2. Industry specific knowledge 3. Combination of multiple data types 4. Robustness 5. Performance requirements 6. Understand failure
Cases Patient Big Data Lab Group, DE eCommerce Product Recom Comp Builder, DE Quality Assurance Chemicals, DE Real Estate Market Prediction Insurer, CH Technology Scouting Defense, CH Bot Communication Wealth Manager, CH Sepsis Diagnosis Medtech, DE Food Recogn App Telco, UK
Merantix GmbH Friedrichstrae 68 DE-10117 Berlin Managing Director: Rasmus Rothe Registration No.: HRB 176845 B Get in touch Rasmus Rothe rasmus@merantix.com Adrian Locher adrian@merantix.com

More Related Content

Deep learning: from theory via virality to applications in business

  • 1. Deep learning: from theory via virality to applications in business Berlin Machine Learning Group Rocket Internet Rasmus Rothe June 2, 2016
  • 9. 100m+ evaluations to date 1m visitors in first 12hrs after launch Spikes in usage due to media attention
  • 10. Its more than a joke Rasmus Rothe, Radu Timofte, and Luc Van Gool. "DEX: Deep EXpectation of apparent age from a single image., ICCVW. 2015. Winner of LAP challenge on apparent age estimation (115 teams) NVIDIA ChaLearn LAP 2015 Best Paper Award Rasmus Rothe, Radu Timofte, and Luc Van Gool. "Some like it hot- visual guidance for preference prediction., CVPR, 2016 Rasmus Rothe, Radu Timofte, and Luc Van Gool. Deep expectation of real and apparent age from a single image without facial landmarks, in submission IMDB-WIKI dataset publicly available: https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/
  • 12. 120,000 profile images 22 million hot-or-not ratings 44.81% mens ratings positive 8.58% womens ratings positive Dataset statistics
  • 13. Datasource # Photos Gender (m/f) Age (23.4yrs) Attractiveness (top 23%)IMDb 500,000 x x Wikipedia 60,000 x x Blinq 120,000 x x x 1) Data to learn from
  • 14. 2) Teaching the deep neural network
  • 15. 3) What happens when you upload a photo
  • 16. There is more research to it ?
  • 17. Prediction RatersVisual space Nearest neighbor Collaborative filtering PT Q 4.0 8.4 2.5 9.7 9.1 3.2 7.7 7.6 6.0 1.5 0.5 8.7 6.0 0.8 2.5 8.5 0.8 8.2 8.9 3.8 8.1 8.2 1.8 7.0 6.7 3.9 7.4 0.8 0.8 5.4 3.9 3.8 8.2 1.8 7.0 6.7 3.9 7.4 5.9 4.0 Visual QueryPastratings OR Matrix factorization Latent space (Q) Regression
  • 18. What led to winning the age estimation challenge (115 teams)? 1. Deeper models (VGG-16) 2. Ensemble classification (20 models) 3. More training data (500k+ images) 5. Robust face alignment 6. Deep expectation layer 4. Augmentation
  • 20. Global Online Players Which of these companies was founded in Germany? Exception:
  • 21. If we look closer 1600out of 1900 of Germanys biggest companies, are B2B Industry Companies. Source: Venohr, B., Fear, J., & Witt, A. (2015). "Best of German Mittelstand- The world market leaders.", Statista V.S. Thats what we are excellent at and lightyears ahead of Silicon Valley. x8 turnover!
  • 22. Merantix is an AI Research Laband a Company Builderthat helps industry companies to leverage the power of AI. Rethinking the German core strengths ENGINEERING
  • 23. Why focus on B2B industries? Industry 1 2 27 201 230 384 B2C Chatbots Gaming Medical Technology Chemicals Machinery & Equipment Automotive Consumer Turner over in EUR in Germany - Sources: Germany Trade and Invest
  • 24. Why applying generic toolboxes is not enough 1. Dirty data 2. Industry specific knowledge 3. Combination of multiple data types 4. Robustness 5. Performance requirements 6. Understand failure
  • 25. Cases Patient Big Data Lab Group, DE eCommerce Product Recom Comp Builder, DE Quality Assurance Chemicals, DE Real Estate Market Prediction Insurer, CH Technology Scouting Defense, CH Bot Communication Wealth Manager, CH Sepsis Diagnosis Medtech, DE Food Recogn App Telco, UK
  • 26. Merantix GmbH Friedrichstrae 68 DE-10117 Berlin Managing Director: Rasmus Rothe Registration No.: HRB 176845 B Get in touch Rasmus Rothe rasmus@merantix.com Adrian Locher adrian@merantix.com

Editor's Notes

  1. intro