Steven T. PIANTADOSI - 3 avril 2025 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)
TITRE : Rules vs. neurons and what may be next
RÉSUMÉ
I will discuss the relationship between large language models and Chomskyan theories of linguistics in the context of the broader debate between rule-based and neural approaches to cognitive modeling. While language models provide a working implementation that surpasses symbolic theories in many respects, I will also present work based in early computer science that seeks to formalize what latent structures must be present in a system in order to generate its observed behavior. This approach is the topic of a forthcoming open textbook, and the approach holds promise for understanding if any grammar-like structures are necessarily present in, for instance, statistical language models. This line of work also points to ways we can rigorously connect neuroscience to behavior.
BIOGRAPHIE
Steven T. PIANTADOSI, Professor in the Psychology Department and Helen Wills Neuroscience Institute, University of California, Berkeley, leads the Computation and Language Lab (CoLaLa). His computational and behavioral research is on the learning of language and concepts, the evolution of human-like cognition, and how ambiguity can serve communicative functions. In his critique of Noam Chomsky’s theory of Universal Grammar, Piantadosi argues that the success of large language models (LLMs) challenges most assumptions of standard linguistic theories.
RÉFÉRENCES:
Piantadosi, S. T., Muller, D. C., Rule, J. S., Kaushik, K., Gorenstein, M., Leib, E. R., & Sanford, E. (2024). Why concepts are (probably) vectors. Trends in Cognitive Sciences, 28(9), 844-856.
Piantadosi, S. T. (2023). Modern language models refute Chomsky’s approach to language. From fieldwork to linguistic theory: A tribute to Dan Everett, 353-414.