Symbol Grounding in AI: A Historical Overview and Recent Neural-Symbolic Approaches

The symbol grounding problem, first articulated by Stevan Harnad and influenced by John Searle’s critique of symbolic AI, raises a fundamental challenge: how can abstract symbols acquire meaning without human interpretation? This talk will explore the historical roots of the problem, tracing its implications in the evolution of AI from classical symbolic reasoning to the rise of statistical and neural approaches. The shift from rule-based systems in the 1990s and 2000s to deep learning and large-scale statistical models has reshaped perspectives on representation and meaning in AI. Finally, we will examine recent neural-symbolic research claiming to solve symbol grounding through neural autoencoders, particularly in structured reasoning tasks such as Sudoku. Through this discussion, we will assess the extent to which these approaches provide a genuine solution to grounding and what challenges remain.