Designing Machine Learning for Humans
summary
Machine learning succeeds or kills based on design details not slogans: a Tempe 2018 AV crash shows radar sees a pedestrian 5.6s out and lidar 5.2s but the vision model was trained only on crosswalks, emergency braking is flagged 1.3s before impact yet is disabled, and policy cut safety riders from two to one, with the human streaming video. Teams therefore run shadow release, perform consequence scanning, test model and human interactions end to end, design for distributed cognition and keep skeptical, diverse pairing and frequent usability checks.
AI-written summary · about
chapters (26)
- 0:00Intro
- 2:23Overview
- 4:01Driving example
- 6:00Car vision
- 7:30A tragic story
- 7:45Tempe Arizona
- 8:26At Night
- 9:10What Happened
- 10:39Who is Responsible
- 15:16Taking Responsibility
- 15:42Autopilot Crash
- 17:16Data
- 18:16Extremes
- 19:52Driver Experience
- 21:56Passenger Experience
- 22:46Team Experience
- 23:58Safety Danger
- 24:50Consequences
- 25:41Testing
- 26:35Working Together
- 27:47Natural Tensions
- 28:30Shared Language
- 30:48Complex Problem
- 31:25Takeaways
- 31:56Conclusion
- 32:46Questions
related talks
44mMachine Learning + Human
43mUX and the Changing Relationship Between Mind and Machine
44mHow to Improve UX with Machine Learning: Insights and learnings until now
41mBeyond the Algorithm: What Makes Machine Learning Work?
48mPrinciples and Practices of Building Machine Learning Systems
47mEthical Machine Learning
04mInterview
45mApplied machine learning: a few lessons I learned the hard way