Stochastic World Models for Verifying Vision-Based Neural Feedback Systems
The paper replaces GAN perception surrogates with stochastic world models built to stay usable for verification.
The authors train the model with physically grounded latent variables and verifier-friendly operations. It matches held-out frames better than GAN surrogates that are up to 130 times larger. Their verification procedure combines falsification, adaptive refinement, symbolic analysis, and backward analysis. On an emergency-braking benchmark, it resolves the full GAN-surrogate state space and more than 80% of the RGB benchmark with the world-model surrogate. ArXiv · AI/CL/LG's note
The authors train the model with physically grounded latent variables and verifier-friendly operations. It matches held-out frames better than GAN surrogates that are up to 130 times larger. Their verification procedure combines falsification, adaptive refinement, symbolic analysis, and backward analysis. On an emergency-braking benchmark, it resolves the full GAN-surrogate state space and more than 80% of the RGB benchmark with the world-model surrogate. ArXiv · AI/CL/LG's note
score 4