PyReason: an AI Logic for Learning

Last week, our group at ASU, Lab V2, released a paper on PyReason on Arxiv. PyReason is a Python package for logical inference and designed for use with machine learning (https://github.com/lab-v2/pyreason). At its core, it uses ideas from Generalized Annotated Logic (Kifer & Subrahmanian, 1992), but integrates key practical extensions for temporal, graphical, and neuro symbolic reasoning. Our idea was to design a logic for use with machine learning.

You may think that’s all fine and good, but are wondering why would we need a logic for machine learning? In this post, I’ll discuss why we did it.

First, a lot of the criticism of machine learning, especially deep learning is that while it obtains excellent result son may tasks, it is merely mimicking historical data and not learning actual relationships. This has resulted in a lot of the major shortcomings in ML such as the hallucinations of large language models, the requirements of vast amounts of training data to learn games, and brittleness in certain applications (e.g., the recent defeat of AlphaGo, difficulty in solving math problems). In a video lecture, we review some of these shortcomings, much of which constitutes active areas of research (part 1, part 2).

Then enter “neuro symbolic” artificial intelligence. Actually an old idea where neural architectures can work hand-in-hand with logic, often even having an equivalence between the two. The idea is symbolic AI has many shortcomings (brittleness to noise, difficulty in learning) that can be address with deep learning while its strengths (modularity, ability to add constraints, symbolic manipulation) can address some of deep learning’s limitations.

Neuro symbolic AI is a highly active area of research, and much of the advancements have identified special logical languages to use in their approach. Our goal with PyReason was to unify many of these logics and provide logic capabilities in a robust and modern Python implementation. We are working on a few joint projects with industry partners applying this to various use-cases, and now we have made the code base and library available as an open source package. In a video, we outline six major capabilities that we felt were important:

1.     Open world reasoning – ability to reason in uncertain situations (important for interfacing with ML models)

2.     Multi-step inference

3.     Explainability

4.     Temporal reasoning

5.     Graph-based reasoning

6.     Designed to support neuro symbolic frameworks

The release of PyReason will kick off not only new research by our group and our collaborators, but also associated software. We’re pretty excited about this new direction!

About the author. Paulo Shakarian, Ph.D. is a tenured Associate Professor at the Fulton Schools of Engineering at Arizona State University. He specializes in artificial intelligence and machine learning – publishing numerous scientific books and papers. Shakarian was named a “KDD Rising Star,” received the Air Force Young Investigator award, received multiple “best paper” awards and has been featured in major news media outlets such as CNN and The Economist. Paulo has been funded by various organizations including IARPA (HAYSTAC, CAUSE, ICARUS), ARO (3x), ONR (5x), and AF/AFOSR (2x). Paulo also co-founded a startup company that used machine learning to predict future exploits; the company was acquired after raising $8 million in venture capital and having obtained over 80 customers. Earlier in his career, Paulo was an officer in the U.S. Army where he served two combat tours in Iraq, earning a Bronze Star and the Army Commendation Medal for Valor. During his military career, Paulo also served as a DARPA Fellow and as an advisor to IARPA. He holds a Ph.D. and M.S. in computer science from the University of Maryland, College Park, and a B.S. in computer science from West Point.