The SwiNGs project (Sampling with Neural Generators), funded by the ANR JCJC program, is launching for the period 2026-2029.
The SWiNGs project aims to harness the potential of generative machine learning to significantly accelerate a fundamental task in scientific computing: sampling. Sampling from a target measure is essential for approximating high-dimensional integrals and plays a crucial role in Bayesian inference and many areas of physics and chemistry. However, the presence of multi-modality in the measure—where likely regions of the state space are separated by unlikely regions—presents a significant challenge for traditional sampling algorithms. These algorithms often become trapped in modes due to their reliance on local exploration. Recent research, including my own, has shown that generative models can address this challenge by facilitating non-local exploration of the state space.
The ambition of the SWiNGs project is to fully realize the potential of generative modeling for sampling by systematically improving the three main components of the implementation pipeline: the models, the sampling algorithms and the training strategies. First, we will enhance the scalability of generative models by incorporating insights from the target measure into their design. Second, we will develop provably exact sampling strategies that leverage generative models, ensuring that errors are controlled and meet the rigorous standards required in scientific computing. Third, we will extend previously introduced integrated training and sampling algorithms to eliminate the need for prior information about the target measure, thereby creating powerful end-to-end sampling pipelines accelerated by generative modeling.