Successful adoption of Machine Learning
1 like122 views
The document outlines the process for successfully adopting machine learning, emphasizing the importance of selecting the right problem to solve and the different types of learning algorithms. It provides case studies illustrating challenges in areas like insurance for young drivers, decentralized energy via solar rooftop, and providing loans to the unbanked, highlighting the data collection and algorithm selection challenges faced. Ultimately, it stresses that users must be incentivized to share data for effective machine learning product development.
Successful adoption of Machine Learning
- 1. Successful Adoption of Machine Learning Rudradeb Mitra | http://www.linkedin.com/in/mitrar/
- 2. Brief Bio • 2002: Published first research paper on AI in an International conference. • 2003-2009: Worked in Germany, Belgium and Scotland at Research Labs, Universities and Startups on AI/ML. • 2010: Graduated from University of Cambridge, UK • 2010-2017: Built 6 startups. • 2017-: Writer. Product Mentor of Google Launchpad. Democratization and Decentralization of building ML products.
- 3. What is Machine Learning? • Learning: Algorithms that can find patterns in past data and predict future patterns. • Three kinds of Learning: Supervised, Unsupervised and Reinforcement.
- 4. How to build successful Machine Learning products?
- 5. Step I • Select the right problem to solve
- 6. How to select the right problem? "Stop identifying cats and start creating value" • Bayesian error (Lowest possible error) rate is >80% • Bayesian error rate is <20%
- 8. Next steps • Selecting the right approach (intuitive or abstract thinking) • Collecting the data (adoption) • Selecting the right algorithm • Building the product (including training and testing the data).
- 9. Three class of problems • Solving problems that were thought unsolvable • Solving problems that were thought not a problem • Improving upon existing systems (error rate >70%)
- 10. Problem 1: Improving upon an existing system Case study: Better risk premiums for young drivers
- 11. • Young drivers have high premiums so insurance companies fight it difficult to attract new customers. The problem
- 13. Next steps • Selecting the right approach: "If we can know how someone is driving then we can calculate better risk" • Collecting the data: How do we get users driving data? • Selecting the right algorithm • Building the product (including training and testing the data)
- 15. Driver’s app Record a trip Trip feedback
- 16. Goals & challenges Rewards 1. Provide incentives
- 17. 2: Cannot force to adopt and let users be in control vs
- 18. • How? 3. Educate your customers
- 19. 4. Create a community
- 20. Results
- 21. What Machine Learning Algorithm to use
- 22. Data but ... • Do not know who is a good or a a bad driver as we do not have labeled data.
- 23. Unsupervised learning Picture taken from: http://www.ai-junkie.com/ann/som/som1.html Find patterns in data
- 24. Problem 2: Problems that were thought unsolvable Case study: Decentralized energy via Solar rooftop
- 25. • Solar adoption is low as the sales process is like 1960s vacuum cleaner sale process. The problem
- 27. Next steps • Selecting the right approach: "If we can know how remotely find rooftops of the people and create a simulator" • Collecting the data: "Use solar satellite images" (public data) • Making the algorithm: "From solar images to calculating rooftop energy potential". • Building the product (including training and testing the data)
- 28. What we want?
- 29. But in reality... In Germany and most of Western world In India
- 30. And google object detection does not work...
- 31. Plus the problem is slightly more complicated with obstacles Water tanks Turbo ventilator Mumpty
- 32. •Type of obstacle in rooftop - We have identified 5-6 types of obstacles. •Edges of the roof - We want to train a machine to learn to identify the edges. •Type of roof Machine Learning to the rescue
- 34. What algorithm to use? Open source code and community!
- 35. Problem 3: Problems that were not a problem Case study: Loans to people without bank account
- 36. • 70% of people in Vietnam don't have a bank account. The problem
- 37. Next steps • Selecting the right approach: "How can we predict future behavior?" • Collecting the data: "Why would users give data?" (because want to get loans) • Making the algorithm • Building the product (including training and testing the data)
- 38. Future behavior of income earnings • Education level • Family background • Current address • Current job and salary
- 39. Unsupervised learning Picture taken from: http://www.ai-junkie.com/ann/som/som1.html Find patterns in data
- 40. Summarizing it all • Select the right problem. • Select the right approach through intuitive thinking. • Collect data via incentivizing users to share data, do not get data behind their backs. • Select the right algorithm(s).
- 41. Key challenge in Machine Learning adoption How do you get data and make users adopt?
- 42. Machine Learning is NOT rocket science Adoption How to collect data? Abstract Thinking Feel free to contact: https://www.linkedin.com/in/mitrar/ mitra.rudradeb@gmail.com Challenges are in Algorithm How to use deal with incompleteness? What data to collect?