.NET Day 19 - The ethical implications and risks of Artificial Intelligence and Deep Learning by Laurent Bugnion

There is no question that Artificial Intelligence and Deep Learning will play an important role in the future (and the present!) of humanity. Taking advantage of faster and faster computers, larger and larger databases, we are able to run very complex algorithms against humongous amounts of data. This allows the creation of tools that can help us in complex areas of our lives. From autonomous vehicles to image and speech recognition, from assisting impaired humans to saving lives in critical situations, from inspecting industrial installations to sending machines in deep space or deep waters, the possibilities are amazing. But this power comes with big responsibilities. How do we take steps to minimize flaws in the data we use for our models? How do we build machines that act for the greater good? What are the risks? In this session, Laurent Bugnion, a Senior Cloud Developer Advocate for Microsoft will talk about what could happen, and what we can do to prevent it.

Transcript

2. 1. First Law: A robot may not injure a human being or, through inaction, allow a human being to come to harm. 2. Second Law: A robot must obey the orders given it by human beings except where such orders would conflict with the First Law. 3. Third Law: A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws.
5. It’s not easy…
6. Algorithm Data Answers
7. Algorithm Data Answers Model
8. Data Answers Predictions New data Model
12. How, can AI be biased? http://gslb.ch/c332s-dotnetday19 2+2 = ? It’s just math!
13. Pneumonia Fact: Asthma patients are less at risk to get Pneumonia How is that even possible?? We can fix it in the data but… WHAT ELSE DID WE MISS?? → Question the whole model
14. Language AI models are often trained with existing lexicons These lexicons are often biased → AI follows our own bias and societal issues → It’s great at “picking up” our flaws “In the span of 15 hours Tay referred to feminism as a "cult" and a "cancer," as well as noting "gender equality = feminism" and "i love feminism now".” http://gslb.ch/c334s-dotnetday19
15. Race bias (from http://gslb.ch/c335s-dotnetday19) “Back in 2015, software engineer Jacky Alciné pointed out that the image recognition algorithms in Google Photos were classifying his black friends as “gorillas.” Google said it was “appalled” at the mistake, apologized to Alciné, and promised to fix the problem.”
16. Race bias http://gslb.ch/c336s-dotnetday19
17. Gender bias Machine learning workforce is 85% male We need emotional intelligence Here too, training datasets are biased “[The AI] would see a picture of a kitchen and more often than not associate it with women, not men,” http://gslb.ch/c344s-dotnetday19
19. AI can be used to “improve” weapon systems AI can be used to spy on people • Face recognition • Classification • Detection of “suspect” behaviors Deep fake
20. http://gslb.ch/c359s-dotnetday19
21. • Respecting the customer choices • Using personal data only as approved • Don’t leak personal data, directly or as inferences • Make sure AI complies with policy and regulations • of the enterprise • of the governments
22. “Millions of people uploaded photos to the Ever app. Then the company used them to develop facial recognition tools. “ Doug Aley, Ever’s CEO, told NBC News that Ever AI does not share the photos or any identifying information about users with its facial recognition customers. Rather, the billions of images are used to instruct an algorithm how to identify faces. Every time Ever users enable facial recognition on their photos to group together images of the same people, Ever’s facial recognition technology learns from the matches and trains itself. That knowledge, in turn, powers the company’s commercial facial recognition products. http://gslb.ch/c358s-dotnetday19
23. • Respect the right of customers to be left alone • Provide them with ways to control how AI handles data • Stop the nagging • If anything, Privacy is becoming more important than in the past AI has to comply
25. It’s not an easy problem to solve • It cannot be solved by maths • No “One solution fits all” • Leaving the hard problems (gender, race etc) from the dataset is also not a solution • Treating the bias adds a new kind of bias  • Think about how the bias affects society
26. The key is in the data • Being fair is important. Don’t exclude one class of population • Humans must remain able to understand the model • Validate assumptions and results • Assume that something will go wrong and keep watching the running system
27. Ethics in the enterprise • Look beyond the next quarterly profit • Enable collaboration • Human insertion benefits AI systems
28. earn trust assist efficiencies dignity transparent accountability
29. http://gslb.ch/c341s-dotnetday19 Webinar: http://gslb.ch/c339s-dotnetday19
31. AI for Good Global Summit The AI for Good Global Summit is THE leading United Nations platform for global and inclusive dialogue on AI. The Summit is hosted each year in Geneva by the ITU in partnership with UN Sister agencies, XPRIZE Foundation and ACM. https://aiforgood.itu.int
32. AI powered captioning http://gslb.ch/c342s-dotnetday19
33. Seeing AI http://gslb.ch/c343s-dotnetday19
34. Our values and principles Enable people Inclusive design Build trust in technology
35. THANK YOU LBugnion@Microsoft.com @LBugnion http://gslb.ch/mycda All you need: http://gslb.ch/dotnetday19
36. Advances in vision technology http://gslb.ch/c338s-dotnetday19