Rich and Expressive Specification of Continuous-Learning Cyber-Physical Systems
Thomas Flinkow, Barak A. Pearlmutter, Rosemary Monahan · 2023
Neural networks are being applied to a growing variety of applications, including safety-critical domains like autonomous vehicles and aircraft, due to their ability to approximate complex functions from limited datasets and adapt by continuing to learn from real-world data after deployment.Ensuring dependability in cyber-physical systems with neural network controllers requires verification beyond linear input-output constraints and local robustness of the network in isolation and must address uncertainties introduced by learning and network behaviour on unseen data, especially when continuous learning after deployment is allowed. Verification of these continuous-learning cyber-physical systems requires assessing probabilistic, temporal, and behavioural specifications that express concepts like resilience and other key properties.Our research seeks to develop strategies and associated tooling for rigorously specifying continuous-learning cyber-physical systems and verifying them against rich and expressive specifications.