Descriptif

« Generative modelling with diffusion : theory and practice »

Generative modelling with diffusion: theory and practice Generative modeling is the task of drawing new samples from an underlying distribution known only via an empirical measure. There exists a myriad of models to tackle this problem with applications in image and speech processing, medical imaging, forecasting and protein modeling to cite a few. Among these methods score-based generative models (or diffusion models) are a new powerful class of generative models that exhibit remarkable empirical performance. They consist of a ``noising'' stage, whereby a diffusion is used to gradually add Gaussian noise to data, and a generative model, which entails a ``denoising'' process defined by approximating the time-reversal of the diffusion. In this talk I discuss three aspects of diffusion models. First, I will present some of their theoretical guarantees with an emphasis on their behavior under the so-called manifold hypothesis. Such theoretical guarantees are non-vacuous and provide insight on the empirical behavior of these models. Then, I will turn to the extension of diffusion models to non Euclidean data. Indeed, classical generative models assume that data is supported on a Euclidean space, i.e. a manifold with flat geometry. In many domains such as robotics, geoscience or protein modeling, data is often naturally described by distributions living on Riemannian manifolds which require new methodologies to be appropriately handled. Finally, I will turn to constraints on the generative process itself. A well-known limitation of diffusion models is that the forward-time stochastic process must be run for a sufficiently long time for the final distribution to be approximately Gaussian. In contrast, solving th e Schröd inger Bridge problem, i.e. an entropy-regularized optimal transport problem on path spaces, yields diffusions which generate samples from the data distribution in finite time. I will present Diffusion Schrödinger Bridge, an original approximation of the Iterative Proportional Fitting procedure to solve the Schrödinger Bridge problem.

Exposé de Valentin De Bortoli (CNRS/ENS) dans le cadre du Data Science Colloquium de l'ENS. 

Thèmes : Informatique
Catégories: Data Science Colloquium
Mot-clés : modélisation, machine learning, intelligence artificielle

Aucun exposé du même auteur.

Can Big Data cure Cancer? Jean-Philippe Vert What physics can tell us about inferenc... Cristopher Moore Beyond stochastic gradient descent for l... Francis Bach The brain as an optimal efficient adapti... Sophie Deneve Cosmostatistics: Tackling Big Data from ... Jean-Luc Starck Searching for interaction networks in pr... Rémi Monasson Brain-computer interfaces: two concurren... Maureen Clerc Machine learning in scientific workflow... Balàzs Kégl Learning Graph Inverse Problems with Neu... Joan Bruna Towards developmental AI Emmanuel Dupoux Optimization's Hidden Gift to Learning: ... Nathan Srebro Prototype-based classifiers and relevanc... Michael Biehl

Valentin De Bortoli
CNRS / ENS / Google DeepMind
Chercheur

Plus sur cet auteur
Voir la fiche de l'auteur

Savoirs ENS
Tous droits réservés
@ 2011 ENS