BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//UNIFR/WEBMASTER//NONSGML v1.0//EN
CALSCALE:GREGORIAN
BEGIN:VEVENT
DTSTART;VALUE=DATE:20250506T171500
DTEND;VALUE=DATE:20250506T171500
UID:17907@agenda.unifr.ch
DESCRIPTION:A large driver contributing to the undeniable success of deep-learning\nmodels is their ability to synthesise task-specific features from data.\nFor a long time, the predominant belief was that 'given enough data, all\nfeatures can be learned.' However, as large language models are hitting\ndiminishing returns in output quality while requiring an ever-increasing\namount of training data and compute, new approaches are required. One\npromising avenue involves focusing more on aspects of modelling, which\ninvolves the development of novel *inductive biases* such as invariances\nthat cannot be readily gleaned from the data. This approach is\nparticularly useful for data sets that model real-world phenomena, as\nwell as applications where data availability is scarce. Given their dual\nnature, geometry and topology provide a rich source of potential\ninductive biases. In this talk, I will present novel advances in\nharnessing multi-scale geometrical-topological characteristics of data.\nA special focus will be given to show how geometry and topology can\nimprove representation learning tasks. Underscoring the generality of\na hybrid geometrical-topological perspective, I will furthermore\nshowcase applications from a diverse set of data domains, including\npoint clouds, graphs, and higher-order combinatorial complexes.\n
SUMMARY:Shapes, Spaces, Simplices, and Structure: Geometry, Topology, and Machine Learning
CATEGORIES:Colloque / Congrès / Forum
LOCATION:PER 08\, auditoire 2.52\, Chemin du Musée 3\, 1700 Fribourg
URL;VALUE=URI:https://agenda.unifr.ch/e/fr/17907
END:VEVENT
END:VCALENDAR