New method models stochastic dynamics without data samples

Researchers have developed a novel method for modeling Schrödinger bridges, which are stochastic dynamical systems that connect two probability distributions. This new approach, termed data-to-energy IPF, allows for the inference of these dynamics even when samples from one or both distributions are unavailable, relying instead on their unnormalized densities. The method is inspired by off-policy reinforcement learning techniques used in training diffusion samplers and has demonstrated success in synthetic scenarios, including learning transports between multimodal distributions. As a byproduct, the technique also offers improvements to existing data-to-data Schrödinger bridge algorithms by learning the diffusion coefficient. The researchers have applied this method to image-to-image translation tasks, enabling data-free translation by sampling posterior distributions in generative model latent spaces. AI

IMPACT Enables new approaches to generative modeling and image-to-image translation by reducing reliance on paired data.

RANK_REASON This is a research paper detailing a new method for modeling stochastic dynamical systems. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• data-to-energy IPF
• Diffusion Models
• Esmeralda Whitammer
• Flow Matching for Generative Modeling
• Iterative proportional fitting
• reinforcement learning
• Schrödinger
• Schrödinger Bridge

• paper
• other

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