Parisa Kordjamshidi (Michigan State University) Hossein Rajabi Faghihi (Michagan State University)

In this talk, I’ll discuss DomiKnowS, our library for Neuro-symbolic programming. It employs a novel declarative framework for designing deep learning models, enabling seamless integration of logical domain knowledge into neural architectures. These architectures can include gigantic black-box language models and other custom pyTorch-based models. The framework represents each learning model as a program, declaring a data model as a graph of concepts and relationships. It allows logical constraints over outputs or latent concepts to be expressed and used in neural models, improving explainability, performance, and generalizability, especially in low-data scenarios. Multiple underlying algorithms have been implemented to integrate logical constraints during training or inference in neural models. The library simplifies programming by separating knowledge representation from learning algorithms. We demonstrate its effectiveness across a wide range of tasks, including natural language processing and computer vision, and plan to extend its application to diverse domains. We will show a brief demo of this framework. The framework is publicly available on GitHub.