Towards Robust, Explainable, and Efficient Foundation Models through Path Integrals, Neuro-Symbolic Reasoning, and Flow-based Computing

ABSTRACT

The sustained widespread adoption of AI into our civilizational fabric needs a principled fusion of human-curated domain knowledge and data-driven neural intelligence, requiring systems that can explain themselves, incorporate human insights, and operate securely under human supervision. Our ability to avoid another AI winter is conditional on the design of AI agents, whereby (1) we can explain the decisions of AI agents in a human-interpretable manner (2) we can communicate our ethics, domain knowledge, and feedback to AI agents using natural languages and their translations into temporal logics, and (3) we can enable human-in-the-loop auditing of the behavior of neural AI agents against regulatory and ethical guidelines using provably correct symbolic AI methods such as model checking and theorem proving. In this talk, I will discuss our ongoing efforts to create solutions — AI models, frameworks, and algorithms that integrate knowledge from natural laws, formal logic, and human judgment with data-driven learning. First, I will introduce new methods for explainability that offer more transparent interpretations of foundation models and traditional deep networks and can also be used to design confidence metrics. Second, I will discuss how probabilistic temporal logic helps translate domain expertise, user intent, ethical guidelines, and human feedback into rigorous machine-interpretable and machine-verifiable directives. Third, I will outline how neuro-symbolic reasoning — combining symbolic AI algorithms, such as model checking and theorem proving, with data-driven foundation models, including large language models such as GPT — can detect inaccuracies in real-world AI outputs and construct provably correct, auditable solutions. I will also share recent progress towards in-memory flow-based computing, revealing how combining hardware efficiency with robust AI can mitigate the soaring energy demands of advanced foundation models and lead to a more sustainable future for AI.

BIOGRAPHY

Dr. Sumit Kumar Jha is an Eminent Scholar Chair Professor of Computer Science at Florida International University in Miami. He earned his Ph.D. in Computer Science from Carnegie Mellon University, followed by several summer faculty appointments with the Air Force Research Laboratory Information Directorate. Dr. Jha has led multi-institutional interdisciplinary teams on projects funded by the National Science Foundation (NSF), Defense Advanced Research Projects Agency (DARPA), Department of Energy (DOE), and other federal agencies, with his research resulting in more than 100 journal articles and full-length peer-reviewed conference publications. This includes publications at highly selective venues, such as AAAI, AISTATS ICLR, IJCAI, DAC, DATE, ICCAD, and NeurIPS. His work has garnered multiple best paper awards (IEEE ICCABS) and nominations at international forums (ACM DAC, ACM/IEEE ICCAD, IEEE MILCOM), as well as the prestigious Air Force Office of Scientific Research Young Investigator Program (AFOSR YIP) Award. Dr. Jha aims to advance trustworthy, responsible, and efficient AI systems that exponentially accelerate innovation in science, engineering, healthcare, and sustainable peace, ultimately transforming the quality of life worldwide.