Bayesian Flow

Bayesian Flow Networks (BFNs) are a class of generative models that iteratively refine the parameters of probability distributions using Bayesian inference, offering a flexible approach to modeling diverse data types (continuous, discrete, and mixed). Current research focuses on improving BFN sampling efficiency through connections to stochastic differential equations and diffusion models, as well as developing novel architectures like semi-autoregressive energy flows for enhanced performance and likelihood-free training. BFNs are proving valuable across various fields, enabling advancements in tasks such as molecule generation, language modeling, and parameter space representation learning, particularly in scenarios with complex or noisy data.