Taming Reachability Analysis of DNN-Controlled Systems via Abstraction-Based Tr...
summary
Claims a simple engineering change: insert an abstraction layer maps reals to intervals so the network treats every interval as indistinguishable, train with those intervals, then run a black-box oracle for abstract states reachability engine that only queries the trained net for actions. The result is a sound, tighter, and faster reachability pipeline that is agnostic to DNN architecture, matches baseline control performance, often verifies benchmarks that white-box tools time out on, and costs modest extra training time; partition explosion is mitigated by allowing linear controllers per cell.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 0:25Research problem overview
- 1:30DNN control systems in safety
- 3:15Reachability analysis
- 5:46Challenges in successor states
- 7:53Abstraction-based training
- 9:18State space partitioning
- 11:08Encoding abstraction layers
- 12:15Experimental results
- 13:21Performance trade-offs
- 14:29Mitigating partition explosion
- 15:13Summary and future work
- 15:55Q&A
related talks
58mBuilding Trust and Safety in Artificial Intelligence with Abstract Interpretation
31mAbstract Neural Networks
18mLearning Abstraction Selection for Bayesian Program Analysis
19mAbstract Neural Networks
56mAbstract Domains for Machine Learning Verification
33mFormal Verification of Probabilistic Deep Reinforcement Learning Policies with(…)
19mInterval Universal Approximation for Neural Networks
21mScaling Abstraction Refinement for Program Analyses in Datalog Using Graph Neural(…)