contact: crek[at]iit.demokritos.gr
This project aims to create a comprehensive knowledge database using First-Order Logic (FOL) extracted from image data. The primary contribution lies in learning predicates along with their arguments, enhancing the interpretability of the extracted knowledge as opposed to relying on anonymous, non-human interpretable symbols. The goal is to advance neurosymbolic AI by grounding FOL representations from raw visual inputs, providing more detailed and human-understandable representations of object interactions, relationships, and the environment. The project builds upon the First-Order State AutoEncoder (FOSAE) model, extending its capabilities to create scalable, semantically rich knowledge graphs that bridge neural perception and symbolic reasoning.
References
Unsupervised Grounding of Plannable First-Order Logic Representation from Images
TV-TREES: Multimodal Entailment Trees for Neuro-Symbolic Video Reasoning