CS5770 Topics in Neurosymbolic Programming
CS5770 Topics in Neurosymbolic Programming
Overview
AI/ML has made remarkable progress over the last decade, with deep learning (neural networks) leading the charge. However, applying these models in safety-critical domains remain challenging due to their black-box nature and lack of interpretability. In contrast, symbolic techniques like symbolic learning, formal methods, and programming languages, have long been the go-to for such domains. Yet, they often suffer from high cost and poor scalability. This has sparked growing interest in combining both paradigms.
Neurosymbolic Programming (see Neurips 22, tutorial for details: https://neurips.cc/virtual/2022/tutorial/55804) is a promising area at the intersection of program synthesis and machine learning. It aims to learn models with program-like structure. One benefit of such programmatic models is that they can naturally incorporate rich inductive biases expressed in symbolic form. Others include modularity, interpretability, and amenability to symbolic analysis.
This advanced reading course explores the fundamentals of neurosymbolic programming and aims to bring students to the state-of-the-art in this emerging area. The course will include lectures and readings covering topics such as neurosymbolic architectures, domain-specific languages, program/architecture search, meta-learning (e.g., library learning), and applications in science and autonomy.
Pre-requisite(s): The course is open to all research students (both PhDs and MTech RAs). Backgrounds(students having done courses) in Compilers/PL/Verification/Synthesis and/or ML-fundamentals/ NLP, are a plus.
Others please drop me an email about your interest and your department and course.
Classroom and Timings
- Instructor: Ashish Mishra
- Where: CS-605
- Discussions Zulip
- When: Slot R, Tuesdays 14:30 - 15:55, Fridays 16:00 - 17:25