Reproduce “Logic Tensor Networks: Deep Learning and Logical Reasoning from Data and Knowledge” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Luciano Serafini, Artur d'Avila Garcez
Logic Tensor Networks (LTNs) offer a unified framework called Real Logic to combine deep learning with first-order logical reasoning. This framework grounds logical symbols in continuous real-valued spaces and enables learning through approximate satisfiability, demonstrated by its ability to perform knowledge completion and infer relationships in an open domain.
Key insights
- Real Logic defines truth values for first-order formulas as continuous real numbers in [0,1], providing a continuous semantic space for logical reasoning.
- Learning is framed as an approximate satisfiability problem, minimizing a 'satisfiability error' over logical clauses, allowing the system to handle inconsistencies in data or knowledge.
- Logic Tensor Networks (LTNs) implement Real Logic by grounding logical symbols (constants, functions, predicates) into learnable real-valued vectors and tensor functions within a deep neural network architecture.
- The framework supports open-domain reasoning by using Skolemization for existential quantifiers, enabling the inference of properties for potentially new or unknown individuals.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.