PhD (M/F): Invariance and transferability in generative models

Descrizione

PhD (M/F): Invariance and transferability in generative models Sintesi in italiano (traduzione automatica): Il candidato sarà inserito nel team Compact di IRISA, uno dei più grandi laboratori di ricerca in Francia nel campo dell'informatica e delle tecnologie dell'informazione. La posizione si concentra sull'esplorazione di modelli generativi che mostrano invarianti di simmetria e trasferibilità di dimensione. Le mansioni principali includono l'analisi teorica delle proprietà necessarie per una buona generalizzazione e la verifica se modelli pratici come le Reti Neurali Grafiche soddisfano tali condizioni. È richiesta una laurea in informatica o un campo correlato. La sede è in Francia e l'arrivo del candidato deve essere autorizzato dall'autorità competente secondo le normative vigenti. The candidate will be attached to IRISA's Compact team. About the lab ============= www.irisa.fr IRISA is one of France's largest research laboratories (over 850 people) in the field of computer science and information technology. Structured into seven scientific departments, IRISA is a laboratory of excellence whose scientific priorities are bioinformatics, systems security, new software architectures, virtual reality, Big Data analysis and artificial intelligence. IRISA is at the heart of the digital transition of society and innovation in the fields of cybersecurity, health, environment and ecology, transport, robotics, energy, culture and artificial intelligence. The position is located in a sector covered by the protection of scientific and technical potential (PPST), and therefore requires, in accordance with regulations, that your arrival be authorized by the competent MESR authority. In recent years, generative models such as diffusion or flow matching have proved to be immensely successful to generate many types of data, from images and videos to text or sound. On the theoretical side, various aspects of these models have been examined, from their generalization to the approximation errors due to various stages of discretization. Naturally, generative models have been defined for many structured data distributions, such as point clouds or graphs. Such data exhibits two crucial, specific properties: 1) they are invariant under various group symmetries (often permutation), and 2) models need to be able to generate samples of different sizes, e.g. training on/generating point clouds with varying number of points. A proper understanding of these properties, both empirically and theoretically, is still crucially missing in the literature. The goal of this PhD is to explore generative models that exhibit symmetry invariances and size transferability. The main focus will be theoretical: we will aim to characterize what properties need to be satisfied by the various ingredients of these models for good generalization and size transferability, and in a second part, examine if models actually used in practice such as Graph Neural Networks do satisfy these conditions. Annuncio in inglese. Fonte: Euraxess (Commissione europea).