Adversarially Robust Abductive Fusion of Pre-trained Transformer-based Perception Models
Leveraging geometric embedding analysis to enhance the reliability of vision transformers in unpredictable environments.
Deploying pre-trained perception models in novel environments degrades their accuracy under distributional shift, and assembling them alone does not recover it: combiners such as majority voting trade recall for precision and are brittle to coordinated failures. Prior metacogniti...
ArXiv CS.AI
Source
Generated insight by OnAirToday AI
This research demonstrates that geometric patterns within model embeddings can replace manual domain knowledge to identify AI errors in novel, unseen environments. By framing perception as an abductive reasoning problem, the system achieves remarkable robustness against adversarial attacks where traditional voting methods fail.
Comments
Keep comments relevant and respectful. No links, advertising, self-promotion, or spam.