New Bayesian inference framework uses transport maps for improved posterior sampling

Researchers have developed a new framework for source-space generalized Bayesian inference that combines efficient few-step prior transports with guarantees for posterior stability. This method represents the prior using an improved MeanFlow (iMF) map and conducts posterior sampling within its Gaussian source space. The framework establishes Wasserstein error bounds between exact and learned posteriors and employs parallel tempering with preconditioned Crank-Nicolson updates, enhanced by a hybrid variant incorporating split Hamiltonian Monte Carlo for improved sampling efficiency. Experiments demonstrate accurate and efficient posterior approximation, with CLIP-guided ImageNet tests showing its capability to steer image priors toward text-specified preferences. AI

IMPACT This research could lead to more efficient and accurate Bayesian inference for tasks involving generative models and test-time guidance.

RANK_REASON Academic paper detailing a new inference framework. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Hamiltonian Monte Carlo
• Hoang Phuc Hau Luu
• ImageNet
• MeanFlow
• parallel tempering
• preconditioned Crank-Nicolson
• Wasserstein

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →