Eric Gaussier
Professor at Université Grenoble Alpes, France
After a PhD, conducted jointly at the scientific center of IBM France in Paris and Université Paris 7, and a research and teaching assistant position at Université Paris 7, I joined the European research center of Xerox in Meylan in 1996. I spent six months at the Palo Alto Research Center (PARC, California) when I was employed by Xerox. I finally joined the Université Joseph Fourier, now Université Grenoble Alpes, in Grenoble as a professor in computer science in 2006. I was Director of the Grenoble Computer Science Laboratory (LIG) from 2016-2020, and am, since Sept. 2019, Director and Scientific Director of the Grenoble Interdisplinary Institute in Artificial Intelligence.
Abstract: Overcoming some of the limitations of modern LLMs (large language models)
Tentative abstract: If LLMs behave very well on in-distribution datasets, they fail to generalize to out-of-distribution datasets, so that they cannot be deployed on data which differ from the datasets they have been trained on. Furthermore, they rely on processes which are costly both in terms of data annotation and computation, so that it is difficult to, e.g., deploy them on long documents. After illustrating some important limitations of LLMs, we’ll present and discuss the solutions that have been envisaged so far.