Neuro-Symbolic Modeling

This course gives an advanced graduate level overview of current methods for unifying neural networks and symbolic logic.

Modern neural network methods have proven to be immensely adept at a variety of domains and tasks. However, these successes require training (or pre-training) on massive amounts of data. Further, despite their impressive performance, deep learning models struggle with issues such as interpretability, explainability and robustness. Towards solving these issues, symbolic approaches can allow us to declaratively specify expected behavior, provide knowledge, and inject reasoning capabilities into AI systems.

In this course, we will explore techniques that combine declaratively stated symbolic knowledge with neural networks at training and prediction time. We will primarily use examples from natural language processing, but also touch upon other sub-fields of artificial intelligence.

See course information for details about course mechanics and policies.

Meetings & Staff

Lectures: Tue & Thu, 9:10AM – 10:30PM.

Instructor: Vivek Srikumar

Discussion forum: We will be using Canvas. Please use the discussion forum as the preferred medium for interacting with the instructor.

Course objectives, or: What can I expect to learn?

By the end of the course, you should be able to critically read current literature and use the ideas learned to:

1. Understand current landscape of techniques for combining symbolic logic and neural networks,

2. Identify or develop learning and prediction paradigms for employing neural networks with declarative knowledge,

3. Be able to design and implement techniques that use symbolic knowledge and declarative constraints to inform machine learning models.