tnreason: Tensor Networks for efficient and explainable Reasoning

The tensor network approach towards efficient and explainable artificial intelligence. Installation The latest version of tnreason can be installed from the Python Package Index (PyPI) using pip: pip install tnreason

Application Demonstrations

• Hybrid Logic and Probabilistic Reasoning: The tensor network formalism generalizes logical and probabilistic reasoning. It therefore enables the combination of hard logical constraints in probabilistic models, which is a form of hybrid reasoning. For a demonstration on hard and soft accounting rules see the Accounting Example.

• Statistical models of Knowledge Graphs: Tensor networks are furthermore useful in storing Knowledge Graphs, and more general of worlds in first-order logic. Based on the sample extraction formalism described in Chapter 11 of the report hybrid logic networks can be trained on data extracted from a knowledge base. For a demonstration of this method on the DPpedia Knowlege Graph see the DBpedia Example.

• Solution of Constraint Satisfaction Problems (CSP): Local constraints can be captured by boolean tensors and CSPs consider contractions of these boolean tensors. Efficient message passing algorithms can be exploited in the solution of these problems. A particular well-known example of a CSP is the game of Sudoku, see the Sudoku Example.

Architecture

For references to the implemented concepts see Appendix A in the report.

References

Tutorials can be found here in the colab demonstrations.

A mathematical report can be found at the documentation repository.