“Certifying Neural Networks for Autonomous Flight,” a Presentation from Daedalean AI
0 likes355 views
The document discusses guidelines for certifying neural networks used in autonomous flight, emphasizing the importance of safety and regulatory compliance. It outlines performance guarantees for machine learning models, including the necessity for independent data distribution across training, validation, and testing sets. The team's expertise in software, aerospace, and regulatory matters supports the development of safe, AI-driven aviation technologies.
“Certifying Neural Networks for Autonomous Flight,” a Presentation from Daedalean AI
- 1. Certifying Neural Networks for Autonomous Flight Edge AI and Vision Alliance — 13 May 2020 David Haber @davhab - dh@ddln.ai
- 3. Non-commercial Commercial Pilot-related 75.3% 63.2% Mechanical 15.0% 25.0% Other/unknown 9.7% 11.8% Fixed-Wing GA Aviation Accidents Source: 26th Joseph T. Nall Report, GA Accidents in 2014, Aopa Air Safety
- 6. No autonomy without AI. But we cannot compromise safety!
- 7. “Concepts of Design Assurance for Neural Networks” Propose a first set of guidelines for NN-based systems facilitating backward & future compatibility with the existing regulatory framework. +
- 10. 10-9 Your system cannot fail in 109 flight hours (on average).
- 13. Performance Guarantees One can mathematically prove that machine learning models generalize.
- 14. First Step Identify input probability space : “All 612 x 512 RGB images you could possibly ever record over Switzerland” Easy to talk about, hard to describe mathematically. Hausen Airport (LSZN), Switzerland
- 16. Show Dtrain, Dval, Dtest are independently distributed in the operating space OS. Data Distribution
- 17. ...not enough! Show Dtrain, Dval, Dtest are independently sampled from the input space and independent from each other. Build an input distribution discriminator Show Dtrain, Dval, Dtest are independently distributed in the operating space OS. Data Distribution
- 18. Algorithm and Model Robustness
- 19. Our Team Strong software/aerospace pedigree: - Computer Vision, Machine Learning, Robotics, Control - Experience with safety critical software development Aviation experience: - Multiple PPL(H)s - Multiple PPL(A)s - Gliding License Focus on regulatory approval: - Dedicated regulatory compliance team - Aviation certification specialist on advisory board We are based in Zürich