Baobab: Compiling OWL 2 DL Ontologies into Differentiable Circuits for Neuro-Symbolic Learning
Recent research published on arXiv reveals Baobab, a novel technique designed to convert OWL 2 DL ontologies into differentiable circuits optimized for neuro-symbolic learning. Addressing the inherent challenges of continuous embeddings, Baobab transforms SROIQ ontologies with a finite ABox into Sentential Decision Diagrams (SDDs). This methodology employs a consequence-based approach to enhance the logical framework, allowing for evidence-conditioned weighted model counting. Experimental results demonstrate that a convolutional neural network can effectively learn to interpret MNIST digits while adhering to specified ontological constraints. The detailed findings are accessible in the paper, uniquely identified as arXiv:2608.17741.
Key facts
- Baobab compiles OWL 2 DL ontologies into Sentential Decision Diagrams (SDDs).
- OWL 2 DL is based on the description logic SROIQ.
- Existing neuro-symbolic learners either embed ontologies in continuous space or restrict to EL++.
- Baobab saturates a propositional core using a consequence-based calculus.
- It instantiates nominals, number restrictions, and role axioms over the active domain.
- The SDD enables evidence-conditioned weighted model counting for training perception networks.
- A CNN learns to read MNIST digits under partial ABox supervision with Baobab.
- The paper is identified as arXiv:2608.17741.
Entities
—