Compositional Approaches to Modelling Language and Concepts

Watch a research lecture from Martha Lewis of the University of Amsterdam exploring compositional and neural approaches to modeling language and concepts, delivered at IPAM's Naturalistic Approaches to Artificial Intelligence Workshop. Examine why large neural language models, despite their effectiveness, still struggle with certain tasks that humans and symbolic systems handle easily. Discover the importance of integrating compositional methods with neural approaches to both enhance model performance and better understand their internal representations. Learn about specific examples where neural models fall short of human capabilities and explore alternative theory-based solutions across language, reasoning, and vision domains. Gain insights into ongoing research directions investigating the reasoning and symbol manipulation capabilities of large neural models. The 57-minute presentation provides a thorough analysis of current challenges and future opportunities in combining different approaches to artificial intelligence.