Symbolic Concept

Symbolic concepts research investigates how abstract ideas are represented and processed within artificial intelligence systems, aiming to bridge the gap between symbolic reasoning and sub-symbolic neural networks. Current efforts focus on identifying and disentangling these concepts within deep neural networks, particularly transformers and large language models, using techniques like high-order derivative analysis and regularization methods informed by symbolic theories. This work is significant for advancing our understanding of both artificial and potentially biological intelligence, offering insights into how complex reasoning emerges from simpler components and potentially leading to more explainable and robust AI systems.