NesySCALE · Scalable and Trustworthy Neuro-Symbolic AI through ε-Guaranteed Approximate Inference

Линиите свързват координатора с партньорите.

Накратко на български

Цел на проекта

Deep learning has achieved remarkable success in tasks such as language generation, image classification, and speech recognition. However, these models remain largely opaque “black boxes,” making it difficult to understand or explain their predictions. This lack of interpretability limits their use in high-risk, regulated domains like healthcare or autonomous systems, where transparency and accountability are essential. The field of neuro-symbolic AI addresses this limitation by combining deep learning’s pattern recognition capabilities with the structured reasoning of symbolic AI. Neural probabilistic programming (NPP) is one such approach, integrating neural networks with probabilistic logic programs to produce models that are both powerful and interpretable. For example, in hospital risk monitoring of sepsis, NPP systems can transform raw medical data into probabilistic observations and then reason reliably using logic programs, making predictions transparent and explainable. A major challenge of NPP is scalability, as symbolic reasoning relies on computationally expensive weighted model counting, which is an obstacle to its application in practice. The NeSySCALE project aims to overcome this by developing a neuro-symbolic framework that integrates learning and reasoning in an end-to-end manner while remaining scalable, robust, and reliable. Using a dynamic depth-first search approach with formal guarantees on approximation errors, the project will enable trustworthy AI in domains where standard neural networks are currently unsuitable. The PF will join the DTAI research group at KU Leuven, under the supervision of Professors Giuseppe Marra and Luc De Raedt.

Оригинален текст от CORDIS (на английски).

Участници

• KATHOLIEKE UNIVERSITEIT LEUVEN · LeuvenКоординаторБелгия

Връзки

• Виж в CORDIS
• DOI: 10.3030/101269988