Monthly Lecture
In collaboration with Sharif University of Technology
This event has taken place

Neuro-Symbolic AI: A Two-Way Connection between Logical Reasoning and Learning
Mohammad Izadi (Sharif University of Technology)
Abstract
Historically, during the first few decades of artificial intelligence research, two broad and largely independent traditions (paradigms) governed the theories of the field and the development of systems based on them: on one side, the connectionist tradition of learning based on statistical methods and artificial neural networks; on the other, the tradition of symbolic AI based on logical reasoning, aiming to automate the inference systems of the various existing logics. Although neural-network systems have achieved very important practical results in learning from data, this tradition has certain shortcomings (such as lack of explainability, difficulty in respecting certain constraints or guaranteeing properties required in critical systems, lack of modularity, no support for multi-step reasoning, and the like). Symbolic AI, on the other hand, is much closer to the human view of notions such as logical reasoning and the explanation of input–output relations and system behaviour, but in practice it is harder to implement and has shown far fewer practical successes. Hence, over the past two decades, combining the two traditions has been proposed, in the hope that their synergy would provide both better explainability and more successful engineering products. This view of AI is called neuro-symbolic, and is sometimes referred to as the third wave of AI.
In this talk, after a quick review of the principles underlying the two traditions, especially from a historical perspective, we discuss some fundamental questions about the relation between the two basic notions of “learning” and “reasoning”: Is intelligence manifested in learning or in reasoning? Do humans (and, in turn, machines) learn to reason? Does reasoning happen in current machine-learning models? We then review some of the most important research of the past two decades that claims to offer neuro-symbolic systems. Some have tried to delegate parts of problems that are in principle about learning from data, or parts of the problem-solving process, to (symbolic) reasoning methods, while others have tried to tackle problems of logical reasoning and inference using learning methods and neural networks.
Date: Monday, 20 October 2025
Time: 16:00–17:30 (Tehran time)
Format: Hybrid (in person and online)
Venue: Sharif University of Technology, Kharazmi Hall, 4th floor, Department of Computer Engineering