Verisig is a tool for verifying properties of neural networks in autonomous systems. The novelty of Verisig lies in its encoding of a deep neural networks as hybrid systems such that it can be easily composed with hybrid systems models of vehicle dynamics and verified using state-of-the-art solvers (e.g., Flow*). Consequently, Verisig has been used to verify safety properties of learning-enabled closed-loop controllers containing neural networks with 10s of layers and 100s of neurons per layer. Verisig 0.9 represents the first public release of the tool, being actively developed as part of the DARPA Assured Autonomy program.
PRECISE releases Verisig v0.9 on GitHub
Sunday, April 14, 2019