报告题目：Data and Symbols: Algorithmic Developments of Neurosymbolic Machine Learning
主 讲 人：Witold Pedrycz 教授
报告时间：2026.7.26 9:00-11:30
报告地点：数理楼224 会议室
报告摘要：
A unified environment of data and symbols has established a promising direction for Machine Learning (ML) by opening a new avenue for learning often referred to as the neurosymbolic paradigm of ML. We advocate that the principles of neurosymbolic learning offer an appealing possibility to explore data and knowledge, with the ultimate goal of achieving efficient learning and addressing the scaling laws of AI. Symbols are central to the representation, elicitation, and processing of knowledge. From the design perspective of ML learning frameworks, data and knowledge are conceptually distinct, as they emerge at different levels of information granularity.
In this talk, we start with a historical perspective on connectivism and symbolism, topics raised by numerous AI researchers including Minsky and Smolensky, among others. We demonstrate the inherent complementarity of these two paradigms. To fully exploit existing ML learning paradigms, we investigate approaches for symbol elicitation and their structural organization into rules or graphs.
The general taxonomy of neurosymbolic constructs involving main categories such as learning-for-reasoning, reasoning-for-learning, reasoning-learning is discussed. We elaborate on the design process guided by a carefully structured additive loss function. Its components minimize discrepancies between the constructed ML model and numerical targets, while ensuring consistency between the model and information granules derived from available domain knowledge. We also illustrate that the symbolic component in neurosymbolic systems provides a unique property known as guardrails for machine learning. Representative neurosymbolic ML architectures, including stable rule-based models and cognitive maps, are presented and discussed.
个人简历：Witold Pedrycz教授，加拿大皇家科学院院士，加拿大工程院院士，波兰科学院外籍院士。现任加拿大阿尔伯塔大学（University of Alberta）计算机智能研究中心主席，国际模糊系统联合会（IFSA）和国际电气电子工程师学会（IEEE）会士（Fellow）。曾任国际模糊系统联合会和北美模糊系统协会主席、国际期刊《Information Sciences》主编，现担任《WIREs Data Mining and Knowledge Discovery》主编，以及IEEE Trans. SMC、IEEE Trans. Fuzzy Systems等多个国际知名期刊编委。长期从事人工智能、模糊系统以及数据挖掘等研究，发表SCI高水平论文900多篇，出版专著15部，为计算智能、粒计算、数据挖掘和不确定性系统建模做出了重要贡献，得到了同行广泛关注和认可。当前谷歌学术引用超11万次，h指数144。
理学院
2026年07月23日
