Evidence Lower Bound

The Evidence Lower Bound (ELBO) is a crucial quantity in variational inference, providing a tractable lower bound on the log-likelihood of complex probabilistic models, primarily used in variational autoencoders (VAEs). Current research focuses on improving ELBO estimation through novel formulations (e.g., incorporating entropy decompositions or using multiple importance sampling), developing more efficient gradient approximation methods, and exploring connections between ELBO and objectives in other generative models like diffusion models. These advancements aim to enhance the performance and interpretability of probabilistic models, leading to improved generative capabilities and more accurate inference in various applications, including image generation and phylogenetic tree inference.