AI-CURE · An Identifiable and Causal Understanding for shortcuts Reduction

Линиите свързват координатора с партньорите.

Накратко на български

Цел на проекта

The AI-CURE project pioneers a new paradigm for tackling shortcut learning in deep neural networks by integrating Causality and identifiability theory—two perspectives that have not yet been unified in this domain. While current strategies against shortcuts rely on heuristics and ad-hoc adjustments, AI-CURE will provide a new theoretical framework that formally explains and mitigates shortcut behavior in neural networks. The project extends recent advances in identifiability from Neuro-Symbolic models by the candidate to general neural representations, and develops a causal theory of distribution shifts to yield provable conditions for robustness. Building on this foundation, AI-CURE will introduce sure-proof mitigation strategies and sanity checks, enabling guarantees for safe model deployment. Finally, by consolidating scattered resources into a unified benchmark, AI-CURE will establish the first systematic platform for evaluating shortcut mitigation. Moreover, it will aggregate different methods and approaches together through a theory-driven field review. This interdisciplinary approach, combining mathematical modelling and AI, not only advances the state of the art theoretically but also provides actionable tools and new community standards, laying the groundwork for safer and more trustworthy AI in critical domains such as autonomous driving.

Оригинален текст от CORDIS (на английски).

Участници

• KOBENHAVNS UNIVERSITET · KOBENHAVNКоординаторДания

Връзки

• Виж в CORDIS
• DOI: 10.3030/101282115