We’re delighted to announce that Tal Linzen will be giving a colloquium this Thurs (Feb 12th) at 11:45 over zoom: https://mcgill.zoom.us/j/86483387308?pwd=2NXkhS1itsS4uYvJIS3tAy3pd9ZE0d.1
All linguistics faculty and graduate students are welcome to attend.
Cognitive science for, and using, large language models
For decades, the fields of artificial intelligence and cognitive science have aimed to create systems that learn and use language like humans. The last few years have witnessed an acceleration in the pace of progress of this endeavor, opening up new avenues to use language models (LLMs) to speed up progress in linguistics and cognitive science on the one hand, and to use cognitive science to advance LLMs and illuminate their capabilities on the other hand. In this talk, I will survey past, ongoing and planned work from my group that pursues these directions. Particular projects I will focus on include training LLMs on synthetic data from formal languages to improve their sample efficiency; using Bayesian models to evaluate and improve LLM assistants’ ability to update their beliefs about users; and using LLMs to understand how humans read syntactically complex sentences. I will conclude by outlining how progress towards LLMs that are more human-like, for example in terms of data efficiency and memory constraints, will advance both cognitive science and interactive AI systems.
Tal Linzen is an Associate Professor of Linguistics and Data Science at New York University and a Staff Research Scientist at Google. He studies the connections between human and artificial intelligence, focusing on language: what can AI systems teach us about humans, and how can we advance AI using what we know about humans? Before NYU, he held positions at Johns Hopkins University and Ecole Normale Supérieure in Paris. His research has been supported by institutions such as Google, the Allen AI Institute, the National Science Foundation, and the National Institutes of Health.