Witold Pedrycz

Data and Symbols: Algorithmic Developments of Neurosymbolic Machine Learning

The unified environment of data and symbols has established a promising direction for machine learning, commonly referred to as neurosymbolic machine learning. This talk explores how the principles of neurosymbolic learning provide new possibilities for integrating data and knowledge, with the ultimate goal of achieving more efficient learning and addressing the scaling challenges of modern artificial intelligence.

Symbols play a central role in the representation, elicitation, and processing of knowledge. From the perspective of machine learning system design, data and knowledge remain conceptually distinct because they emerge at different levels of information granularity. The talk will revisit the historical relationship between connectionism and symbolism, including ideas developed by researchers such as Minsky and Smolensky, and demonstrate the inherent complementarity of these two perspectives.

Particular attention will be given to the elicitation of symbols and their structuralization into rules and graphs. A general taxonomy of neurosymbolic constructs — including learning-for-reasoning, reasoning-for-learning, and reasoning-learning — will be discussed. The presentation will further examine the design of additive loss functions that combine numerical learning objectives with adherence to information granules representing available knowledge.

The symbolic component of neurosymbolic systems will also be considered as a mechanism for machine-learning guard railing. Representative neurosymbolic architectures, including stable rule-based models and cognitive maps, will be presented to illustrate these concepts.