Date and Time: 29/10/26, 5pm CET
Speaker: Marco Baroni, Departament de Traducció i Ciències del Llenguatge, Universitat Pompeu Fabra (Spain)
Presenter: TBA
Abstract:
Recent studies suggest that, as their performance improves, vision-language models based on deep neural networks converge towards shared representations across different architectures, modalities, and languages. This indicates that they all develop a “semantic core” in which they encode abstract representations of input meaning. In this talk, I will present work in which we attempt to better characterise this semantic core: where it emerges within the models, how it is affected by language and modality, and to what extent it is possible to disentangle purely semantic representations from language-specific ones, such as syntactic representations.
Bio: Marco Baroni is an ICREA Research Professor at Universitat Pompeu Fabra and holds a PhD in Linguistics from UCLA. He previously worked at the Centre for Mind/Brain Sciences at the University of Trento and at Facebook AI Research (FAIR) in Paris. His research on multimodal and compositional distributional semantics has received major accolades, including two ERC grants and the ACL Test-of-Time Award. He currently studies the internal workings of large language models and what they can teach us about how humans represent language.
Enlace a la charla: https://zoom.us/webinar/register/WN_4K022tk2QHyjYDtOo3zviQ
