Perception Contracts for Safety of ML-Enabled Systems
summary
Proposes Perception contracts — symbolic formulas that bound error between ground truth and perception estimates so a controller keeps a safety invariant — and a synthesis pipeline called Perceptor learner using an exact decision‑tree learner, quantifier‑free octagonal constraints, and Gurobi as the teacher/solver. Implements and tests on a LaneNet model (18 layers, >3M neurons, ~24,000 GT/GTE pairs) and a ResNet‑18 crop‑follow system (~7,000 pairs). Contracts synthesize effectively, reveal NN perception breaks safety under real inputs, and work well for runtime monitoring.
AI-written summary · about
chapters (14)
- 0:00Perception contracts intro
- 1:29Safety for ML cyberphysical systems
- 2:32Limitations of traditional contracts
- 3:37Perception contracts insights
- 5:27Formal definition of contracts
- 7:57Synthesis problem for contracts
- 8:39Learning algorithm for synthesis
- 10:25Implementation tools
- 10:48Case studies evaluation
- 12:14Safety under imperfect perception
- 13:16Invariants for lane tracking
- 14:42Results effectiveness
- 16:50Runtime monitoring
- 18:33Q&A
related talks
49mUsing Deep Learning for Perception in Autonomous Systems: A Perspective from Control Theory
01h 10mNeurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)
08mUnified Perception: Models for Autonomous Driving
55mTowards Verified Deep Learning
01h 35mDeep Robotic Learning
18mCustomizable Computer Vision Expands Data Access Without Compromising Privacy
38mTowards Privacy-Respectful Cyber-Physical Systems: Two Perspectives
01h 07mTheory of Computation II