Architecting IoT with Machine Learning
•
0 likes•354 views
The document discusses how machine learning can help architect Internet of Things (IoT) systems for widespread consumer adoption. It describes three examples of using machine learning with IoT data: (1) identifying patterns of risky drivers to adjust insurance premiums, (2) predicting short-term driving behavior to improve road safety, and (3) using long-term driving history with recurrent neural networks to provide customized nudging to change driver behavior over time. The document argues that machine learning can create value from IoT data and benefit consumers by making systems safer, lowering costs, and incentivizing good behaviors.
Report
Share
More Related Content
What's hot
What's hot (20)
Similar to Architecting IoT with Machine Learning
Similar to Architecting IoT with Machine Learning (20)
More from Rudradeb Mitra
More from Rudradeb Mitra (12)
Recently uploaded
Recently uploaded (20)
Architecting IoT with Machine Learning
- 1. Architecting IoT systems with Machine Learning Rudradeb Mitra | https://www.linkedin.com/in/mitrar/ Data Science Summer Conference
- 2. The vision of IoT by Intel https://www.youtube.com/watch?v=rnDey89wp_M&t=31s
- 3. How many of you live in this home?
- 5. What do I want talk about? • How to build consumer adopted IoT products? • Through stories of my experiences • Some problems you can solve!
- 6. My 15 years on and off affair with IoT connecting the dots...
- 7. 2003, Bremen, Germany RoboCup Japan 2003, Bremen, Germany
- 8. To solar adoption through satellite images
- 9. The story of IoT from 2003-2018 IoT
- 10. So what went wrong...
- 11. Consumer IoT systems Database View data
- 12. What are the problems? • Viewing past data has almost no value • Most IoT startups ended up just collecting big data • Very few killer applications which creates x5-x10 value
- 13. Hidden cost of IoT Relational Data Time series 'Big' Data (activity, time)(name, address)
- 14. Relational legacy data Big data (eg time-series data) Database eg. Postgres Database eg. Mongo, Casandra Application logic (to split and combine data into relational / non-relational part) User Query SQL query Time series data query Database architecture of IoT @copyright: Rudradeb Mitra
- 15. Data Volume and Speed @copyright: Rudradeb Mitra
- 16. For adoption technology needs to cross the chasm • You have to solve a real problem for them. Build painkillers, not vitamins!
- 17. So how to cross the chasm?
- 18. Let me tell you a story...story from one of my startup
- 19. Driver’s app Record a trip Trip feedback Goals & challenges Rewards
- 20. Results Over 4Mkm driving data
- 21. IoT Architecture Database Engagement Add a switch to turn off/onView Trip • Incentivizing • Nudging • Not forced
- 22. What is missing? Do you think majority will adopt?
- 24. When there is nothing else, Machine Learning comes to rescue!
- 25. What is Machine Learning? • "learn patterns from data without explicitly being programmed"
- 26. Why is Machine Learning useful?
- 28. IoT Architecture with Machine Learning Database Gamification Social Engagement Add a switch to turn off/onView Trip Predictive + Analytics Machine Learning Algo @copyright: Rudradeb Mitra
- 29. What are the values? • Find risky drivers and adjust their premiums so that majoity pay less • Predict short and long term future to make roads safer by changing behaviors
- 30. Lets see using Machine Learning (esp. Neural Nets)
- 31. 1. Can we find patterns of risky drivers?
- 32. Patterns of risky drivers - Clustering Picture taken from: http://www.ai-junkie.com/ann/som/som1.html Find patterns in data
- 33. Self Organizing Maps (Clustering) 0.1 0.3 - Euclidean Distance between training vector and weights - The best match node is selected - Adjust weights of neighboring nodes to match this weight 0.55 0.15 0.31 0.49 Training Data 0.6 0.9 0.8 @copyright: Rudradeb Mitra
- 34. Picture taken from http://www.ai-junkie.com/ann/som/som1.html Most risky drivers Most safe drivers For the driving example
- 35. And how is that helpful? • Premiums can be adjusted according to driving behavior and thus majority will end up paying less
- 36. 2. Can we go further and predict short term future?
- 37. word2vec- Predicting next word word2vec - Does not understand words or grammar How Feeling You Are Today Input words embeddings Predicted next word Hidden Layer Output word embedding
- 39. Example word xi N= 2 0 0 0 1 0 0.3 0 0.7 0 ... 0 0.1 0.77 0.39 . . 0 0.29 0 0.55 0 ...0.3 0.5 0 Wi = Wo = Word embedding for Xi = X. Wi . Wo Word embedding Xi = [0.33 ... 0.64 ] 0 0 0 1 X =
- 40. word2vec - 2d space - Output matrix to 2 dimension using Principal component Analysis
- 41. Picture taken from https://www.lucypark.kr/courses/2015-ba/text-mining.html Germany - France + Paris = Berlin
- 42. Prediction for drivers score acceleration on Friday score braking on Friday word2vec - Using the semantic association between orders (Product ID) (Predicted score on Sunday) score braking on Saturday score speed on Saturday
- 43. And how is that helpful? • Advance short term warnings making roads safer and saving lifes
- 44. 3. And what about long term future?
- 46. Long Short Term Memory Cell state Picture taken from http://colah.github.io/posts/2015-08-Understanding-LSTMs/
- 47. NN NN NN f NN Forget gate Input gate Output gate New information Forget an old information Updated information added @copyright: Rudradeb Mitra
- 48. LSTM cell LSTM cell LSTM cell Score Score Score Driving history Time 1 Time 2 Time 3 @copyright: Rudradeb Mitra
- 49. And how is that helpful? • Provide nudging techniques to change behaviors
- 50. To summarize • Roads safer • Adjust premiums saving money • Incentive to good behaviors
- 51. Who will benefit • Majority of people like you and me, who are GOOD drivers!
- 52. What other applications are out there?
- 53. 03/06/2010 2010, Prague, Czech Republic
- 54. What can we achieve? • Machine Learning can predict your future consumption (short term, long term). • If I know how much I will consume, I can save money and save electricity!
- 55. Who will benefit to people who have been left behind...
- 57. Output Fully connected pooled feature maps pooled feature maps feature maps feature maps Input image 32x32 Convolutions 6@28x28 Subsampling Convolutions Subsampling 6@14x14 16@10x10 16@5x5 Convolutional Neural Networks
- 58. Who will benefit to people who have been left behind...
- 59. Healthcare Database Gamification Add a switch to turn off Not to punish, social encouragement Learning Algorithm @copyright: Rudradeb Mitra
- 60. What can we achieve? • If I know advance health warning, I can save lives! • Medical cost can do down
- 61. Retail
- 62. Future
- 64. IoT needs Machine Learning for adoption Feel free to contact: https://www.linkedin.com/in/mitrar/ mitra.rudradeb@gmail.com