Reproduce “Bayesian Learning via Neural Schrödinger-Föllmer Flows” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Francisco Vargas, Andrius Ovsianas, David Fernandes, Mark Girolami, Neil D. Lawrence, Nikolas Nüsken
Neural Schr¨odinger-Föllmer Samplers (N-SFS), a stochastic control framework for approximate Bayesian inference, achieves competitive predictive performance and improved uncertainty quantification compared to Stochastic Gradient Langevin Dynamics (SGLD) across various tasks. The method demonstrates an order-of-magnitude reduction in Expected Calibration Error (ECE) on CIFAR10 and effectively handles complex, multimodal posteriors in high-dimensional models like hyperspectral image unmixing.
Key insights
- Stochastic control offers a framework to reach target posterior distributions in a finite, bounded time, providing a theoretical advantage over MCMC methods that rely on asymptotic convergence.
- Neural network parameterization enables N-SFS to flexibly approximate complex drift dynamics and scale to high-dimensional Bayesian Deep Learning tasks through amortized inference.
- The 'sticking the landing' estimator effectively reduces gradient variance, leading to more stable and efficient training, and the N-SFS objective can be connected to the negative Evidence Lower Bound (ELBO) in variational inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.