Beyond Maximum Entropy: A new paradigm for Modeling, Inference, and Learning Efficiency

Description du projet

Simplifier la complexité de l’apprentissage automatique

Les modèles d’apprentissage automatique deviennent de plus en plus complexes afin de faire progresser des domaines tels que les soins de santé, la finance et les sciences du climat. Cependant, les modèles actuels sont gourmands en ressources, difficiles d’accès et peu interprétables, ce qui limite leur potentiel. Le projet BeME, financé par le CER, s’attachera à développer des modèles plus simples et plus efficaces, qui resteront performants tout en étant faciles à comprendre. En recourant à des techniques issues de la physique statistique, de l’étude des systèmes désordonnés et de la physique numérique, les chercheurs créeront des outils capables de révéler des tendances dans les données, de fonctionner efficacement même avec des informations limitées et de réduire les coûts d’entraînement. Des applications en neurosciences, en bio-informatique et en modélisation de la turbulence démontreront l’incidence concrète de ces innovations. Les travaux proposés visent à remodeler l’avenir de l’apprentissage automatique afin de le rendre plus accessible et plus efficace face à un éventail plus large de défis.

Objectif

Machine learning is revolutionizing science and society, enabling transformative breakthroughs such as predicting protein structures, simulating many body quantum systems, and redefining entire fields like healthcare, finance, or climate modeling. Yet, this rapid progress comes at a cost: the unprecedented scale and complexity of state-of-the-art models make them resource-intensive, inaccessible, and often opaque. These challenges limit their potential, hindering our ability to understand, optimize, and deploy them responsibly. To unlock the full power of machine learning, we must reimagine its foundations.
This project challenges the 'bigger is better' paradigm in machine learning by proposing an alternative: the development of simpler, more efficient generative models that can handle the complexity of the real world while remaining analytically interpretable. Using techniques from statistical physics, disordered systems and computational physics, the project aims to uncover how these models encode patterns, address learning challenges and reduce training costs without sacrificing performance or expressiveness. Crucially, these models will be designed to extract meaningful insights from general-purpose data sets, perform reliably in data-scarce scenarios and accelerate simulations in complex systems.
By developing a new generation of inference algorithms and showcasing their effectiveness through three proof-of-concept applications in neuroscience, bioinformatics, and turbulence modeling, this research aims to significantly expand the toolbox for data-driven discovery. Beyond tackling practical challenges, it seeks to advance our fundamental understanding of unsupervised learning mechanisms while prioritizing sustainability, transparency, and inclusivity. This project reimagines machine learning as a responsible and accessible tool to drive scientific and societal progress.