Hidden Beauty Behind Generative AI: Variational Inference and ELBO

Explore the mathematical foundations of generative AI in this 32-minute video that delves into Variational Inference and Evidence Lower Bound (ELBO). Learn how these crucial concepts enable intelligent systems to model probability distributions in practical ways. The presentation covers the problem setup, latent variable formalism, distribution parametrization, training objectives, importance sampling, variational distribution, and a detailed explanation of ELBO. Presented by Artem Kirsanov, a graduate student at NYU Center for Neural Science and researcher at Flatiron Institute, this educational video provides a comprehensive look at the mathematical beauty underlying modern AI systems. The content follows a structured approach with clear sections from introduction to conclusion, making complex concepts accessible.