Webinar: Introduction to Neuro-Symbolic AI
For AI/ML engineers, technical architects, and applied researchers evaluating whether foundation models alone are enough for high-stakes reasoning tasks — and for anyone building AI into legal, financial, medical, or other domains where "mostly right" isn't good enough.
Key Takeaways
- Scale solves perception, not reasoning. More data and compute keep making models better at recognizing patterns in noisy, real-world input — but genuinely novel situations, ones a model hasn't effectively seen before, require actual reasoning, and that's exactly where today's models still fail. (Dr. Parisa Kordjamshidi)
- What looks like LLM "reasoning" may really be sophisticated pattern-matching. Rather than working step-by-step through premises to a conclusion the way formal logic does, an LLM often recognizes that a case resembles something it's seen before and produces the associated answer. That's fast and often correct — but when a pattern hasn't been sufficiently observed, the same mechanism can produce confident, plausible-sounding, and simply wrong answers. (Dr. Parisa Kordjamshidi)
- Neurosymbolic AI is a spectrum, not one architecture. On one end: loosely-coupled systems where a symbolic tool (a theorem prover or model checker) acts as an external guardrail on an LLM's output — for example, checking that a credit decision obeys required rules before it's finalized. On the other end: fully integrated systems where the neural network's own architecture is built to natively learn and perform sound reasoning rules, with no separate symbolic layer at all. (Luis Lamb)
- The best scientific breakthroughs in history were rarely "data-first." Darwin, Newton, Einstein, and Maxwell all built new explanatory theories and structures first, then used existing data to validate them afterward — not the reverse. Mendeleev's periodic table is one of the few genuinely data-first breakthroughs. This is central to why symbolic, reasoning-first approaches remain necessary even as models and datasets keep growing. (Luis Lamb)
- A concrete failure mode: spatial reasoning. Because LLMs are trained on internet-scale data rather than embodied, real-world interaction, they often lack basic spatial concepts — like "frame of reference" (whether "left" means left of the camera, or left from a specific person's own perspective) — and struggle badly with compositional descriptions that combine several simple concepts in combinations they haven't specifically encountered, like "the woman in the red shirt to the left of the car." (Dr. Parisa Kordjamshidi)
- A practical open-source framework for combining the two: Dr. Kordjamshidi's team built a Python library (referenced in the discussion as "Domino") that lets developers declare concepts and relationships as a graph, connect each concept to any neural model — from a simple perceptron to a full LLM — to get a probability, and then combine those probabilistic judgments through logical rules or constrained optimization (such as integer linear programming) to reach a final answer that respects hard constraints and stays explainable. (Dr. Parisa Kordjamshidi)
- Symbolic reasoning matters most exactly where "95% right" isn't good enough. Legal, medical, and financial domains, safety-critical code verification, and any use case involving logical negation are areas where getting from 95–98% accuracy to genuinely reliable (100%) is a fundamentally different — and much harder — bar than typical benchmark scores suggest. (Luis Lamb)
- RAG alone doesn't guarantee reliable reasoning, even when retrieval works perfectly. In legal-reasoning research, Dr. Kordjamshidi's team found that even when RAG successfully retrieves the correct, relevant context, that doesn't mean the model will reliably reason through it correctly. Perturbation testing revealed models that scored well on standard benchmarks were actually quite brittle — small changes to the input exposed that the model wasn't genuinely following the logical chain at all. (Dr. Parisa Kordjamshidi)
Q&A Highlights
The following questions were raised by the panel moderator during the live session.
There's a narrative that foundation models are getting good enough at reasoning that symbolic systems may become unnecessary. Where does the real boundary sit today?
Luis Lamb pushed back on the premise directly: history's major scientific leaps weren't data-first discoveries. Darwin, Newton, Einstein, and Maxwell each applied a new explanatory theory to data that already existed, then used the data to validate the theory afterward — not the other way around. Today's AI systems can meaningfully accelerate data-first discovery, but that's a different capability than genuine reasoning. Dr. Kordjamshidi agreed, adding that verification and reliability are inherently symbolic-side concerns: no amount of scale removes the need for rigorous mechanisms to check a model's reasoning, particularly for genuinely new situations it hasn't encountered before.
Are there concrete examples where neurosymbolic AI solved something a pure LLM approach couldn't?
Luis Lamb pointed to several: rule-bound financial and legal applications where rule violations simply aren't tolerable, real cases where LLM-generated outputs violated required rules with real downstream consequences, and — a deeper theoretical issue — LLMs' persistent difficulty handling logical negation, a well-studied hard problem in symbolic logic that he suggested many LLM developers may not fully appreciate the complexity of. He also flagged safety-critical code and system verification: reaching 90–95% correctness is comparatively easy, but many domains require closing the gap from 95–98% to 100% — a fundamentally different bar than typical benchmark accuracy represents.
If a company has already invested in RAG, is it worth adding a symbolic layer on top — or is better prompt engineering and retrieval enough?
Dr. Kordjamshidi said yes, based directly on her team's legal-reasoning research: even when RAG successfully retrieves the correct, relevant information into an LLM's context, that doesn't guarantee the model will reliably reason through the necessary logical steps to reach a correct conclusion. Perturbation testing showed models that looked highly accurate on standard benchmarks were actually quite brittle underneath — small changes to the input revealed the model wasn't genuinely following the reasoning chain at all. For sensitive domains like legal and medical applications, she argued an explicit symbolic reasoning mechanism is still necessary on top of RAG, not a substitute for it.
Panelists
Dr. Parisa Kordjamshidi
Associate Professor, Computer Science and Engineering, Michigan State University, DaaX Advisor
Dr. Kordjamshidi, DaaX Advisor, is an Associate Professor in the Department of Computer Science and Engineering at Michigan State University (MSU), with an affiliation as a research scientist at the Florida Institute for Human and Machine Cognition. Prior to MSU, she was an assistant professor at Tulane University (2016–2019) and a postdoctoral researcher at the University of Illinois at Urbana-Champaign (2013–2016). She obtained her PhD from KU Leuven, Belgium, in July 2013. Her research interests include machine learning, natural language processing, combining learning and reasoning paradigms, and learning-based programming. She is a recipient of a National Science Foundation CAREER Award (2019).
Luis Lamb
Board Member, Centaur AI Institute, Special Advisor for AI Engagement, Stony Brook University
Luis Lamb is a global expert in artificial intelligence (AI), innovation, and technology strategy, with leadership experience across academia, government, and industry. He holds a PhD in Computer Science from Imperial College London and an MBA from the MIT Sloan Fellows program. He is a pioneer in Neuro-Symbolic AI, co-authoring Neuro-Symbolic Cognitive Reasoning, a landmark in the field. He has built and led AI and machine learning teams in corporate and research environments, driving value creation, productivity, and robust AI/MLOps practices. His expertise spans AI applications, innovation strategy, technology management, and entrepreneurship.
Sajjad Khazipura
Co-Founder & CTO, DaaX
Co-Founder & CTO of DaaX. At DaaX, Sajjad is building LAKEer — an enterprise neuro-symbolic AI platform combining knowledge graphs, vector search, and structured data retrieval to deliver grounded, hallucination-resistant AI for mission-critical applications. He brings experience building internet-scale, high-reliability systems, petascale data platforms, real-time transactional systems, and industry AI solutions across e-commerce, manufacturing, healthcare, and finance. Prior roles include Principal Solutions Architect at AWS and VP & Global Practice Head at Wipro Technologies.
Next webinar
- Webinar: AI in Oil & Gas: Turning Complex Data into Actionable Intelligence
- Continue the DaaX webinar series with a practical look at AI in oil & gas operations.