PyReason

What is neuro-symbolic logic?

Neural networks learn patterns from data. They are strong at perception, such as reading text or recognising an image, but the reasons behind their answers are hard to inspect, and they cannot be relied on to follow an explicit rule.

Symbolic logic works the other way round. It reasons with explicit facts and rules that people can read, and every conclusion can be traced back to the rules that produced it. On its own, though, it cannot learn from raw data.

Neuro-symbolic AI combines the two. A neural model turns raw data into facts, often with a measure of confidence, and a logic engine reasons over those facts with readable rules. The result learns from data and can still explain itself.

What PyReason is

PyReason™ is open-source Python software for explainable reasoning over graphs, built on annotated, temporal and open-world logic. It is developed by Lab V2 at Syracuse University and released under the BSD-2-Clause licence.

• Graph-based reasoning

  Applies logical rules to knowledge graphs and other networks, loaded from GraphML or NetworkX.

• Annotated logic

  Every fact carries a confidence interval between 0 and 1 rather than a plain true or false: [1, 1] is true, [0, 0] is false and [0, 1] is unknown.

• Open-world reasoning

  Anything not stated is treated as unknown, not as false.

• Temporal rules

  Rules can say "if A holds, then B holds after n time steps".

• Explainable by design

  It keeps a trace of every inference step behind each conclusion.

• Works with machine learning

  Predictions from a PyTorch classifier can be added as facts and reasoned over with rules. This is where PyReason is neuro-symbolic. See the machine-learning integration guide.

pip install pyreason

PyReason™ is a trademark of the Arizona Board of Regents / Arizona State University. Rational Agents is not affiliated with Lab V2.

How to support PyReason

• Star it on GitHub

  Stars help other developers find the project.

  lab-v2/pyreason on GitHub

• Try it and read the docs

  Install the package from PyPI and work through the documentation.

  DocumentationPackage on PyPI

• Report bugs and request features

  Open an issue when something breaks or a feature is missing.

  GitHub issues

• Ask questions and share what you built

  GitHub Discussions is the place for questions, ideas and projects made with PyReason.

  GitHub Discussions

• Contribute code

  Start with the contributing guide: pre-commit hooks run the unit tests on every commit and the functional tests on every push.

  Contributing guide

• Cite the paper

  If PyReason helps your research, cite the AAAI Spring Symposium paper.

  arXiv:2302.13482

  @inproceedings{aditya_pyreason_2023,
  title = {{PyReason}: Software for Open World Temporal Logic},
  booktitle = {{AAAI} Spring Symposium},
  author = {Aditya, Dyuman and Mukherji, Kaustuv and Balasubramanian, Srikar and Chaudhary, Abhiraj and Shakarian, Paulo},
  year = {2023}}

• Explore related work

  PyReason Gym: PyReason in a grid-world reinforcement-learning setting.

  PyReason Gym