Efficient Diffusion Models for Scientific Machine Learning




Overview of multi-stage diffusion model strategy for improving training and sampling efficiency.

Recently, diffusion models have emerged as a powerful new family of deep generative models with record- breaking performance in many applications. However, due to the intensive requirements of data and computational resources, diffusion models suffer from limitations that deter their practical use in many scientific applications such as medical imaging. The goal of this project is to develop computationally and data efficient generative diffusion models to satisfy the unmet demands of real-world scientific and medical applications. Specifically, our research team propose novel techniques to improve the training and sampling efficiency of generic diffusion models, and also delve computationally efficient diffusion models in latent space for high-dimensional data to further enhance data, memory and time efficiency. The computationally efficient techniques developed in this project will be crucial to open the door to many critical and challenging scientific applications such as high-dimensional medical imaging.

Other Researchers

Liyue Shen (Electrical Engineering and Computer Science)

Jeff Fessler (Electrical Engineering and Computer Science)

Qing Qu (Electrical Engineering and Computer Science)

Mark your calendar for the MICDE SciFM 2024 Conference on April 2nd & 3rd, 2024!