“Understanding Diffusion Model Objectives from the Perspective of ELBO”
Ruiqi Gao, Research Scientist, Google DeepMind
To achieve the highest perceptual quality, state-of-the-art diffusion models are optimized with weighted training objectives that look very different from the maximum likelihood and the Evidence Lower Bound (ELBO) objectives. In this talk, I’ll revisit the connection between diffusion model objectives and ELBOs, and show that they are actually closely related.
Symposium overview: Generative AI: Diffusion Models for Scientific Machine Learning mini-symposium, co-organized with the Michigan Institute for Data Science and the department of Electrical and Computer Engineering at the University of Michigan.
Recently, diffusion models have emerged as a powerful new family of deep generative models with record-breaking performance in many applications, including image generation, audio synthesis, inverse problem solving, and many scientific disciplines. However, despite their impressive results, they also encounter numerous challenges and constraints that inhibit their practical implementation in many scientific pursuits. Therefore, this mini-symposium, co-organized by MIDAS and the Department of Electrical and Computer Engineering, will serve as a timely platform where experts and researchers from both methodology and application research fields will explore the latest progress and developments in generative AI and diffusion models, and delve into the application of these models in scientific and medical fields, which will become a prime venue for idea exchange and fostering research partnerships in this emerging field.
For more information about the series and speakers, please visit the event page at: [ Ссылка ]
- Sign up for the MIDAS newsletter: [ Ссылка ]
- Twitter: @um_midas
- LinkedIn: [ Ссылка ]
#datascience #artificialintelligence #research #ai #generativeai
