2024-25 Distinguished Computational Linguistic Lecture
Finding linguistic structure in large language models
with Christopher Potts, Ph.D.
Professor and Chair in the Department of Linguistics at Stanford University
Neural network interpretability research has proceeded at an
incredible pace in recent years, leading to many powerful techniques
for understanding how large language models (LLMs) process
information and solve hard generalization tasks. In this talk, I'll argue
that these same techniques can provide rich evidence for linguistic
investigations. My focus will be on the framework of causal
abstraction, a family of interpretability techniques that allow us to test
sophisticated hypothesis about the structures latent in LLM
representations. I'll argue that analyses in this mode can directly
inform deep analyses of specific linguistic phenomena, and that they
can yield insights into the kinds of inductive bias that are sufficient for
acquiring natural languages.
Event Snapshot
When and Where
Who
Open to the Public
Interpreter Requested?
No