Stochastic World Models for Verifying Vision-Based Neural Feedback Systems
Researchers verify vision-based neural controllers using compact stochastic world models that outperform much larger GAN surrogates.
The paper proposes stochastic world models with physically grounded latents as perception surrogates for verifying vision-based neural feedback systems. These models reproduce held-out frames more faithfully than GAN surrogates that have up to 130 times as many parameters. Combining falsification, adaptive refinement, symbolic, and backward analyses resolves the entire emergency-braking state space, including 38% left unresolved by a prior verifier, and over 80% of an RGB benchmark.
- World models beat GAN surrogates with up to 130 times more parameters.
- Procedure resolves the full emergency-braking state space on a GAN surrogate.
- RGB benchmark: more than 80% of the state space is resolved.
From the original article
1 min read at arxiv.orgVerifying a vision-based neural feedback system requires a model of the observations its controller acts upon. Such a model must capture the variation the sensor produces, while remaining tractable for closed-loop analysis. Generative adversarial networks (GANs) have served as perception surrogates, but they are large, reproduce complex scenes poorly, and are hard to verify.
By I. Samuel Akinwande · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.