|
3:30 p.m.
Zoom online presentation
For More Information:
Soheil Feizi
sfeizi@umd.edu
https://umd.zoom.us/j/95197245230?pwd=cDRlVWRVeXBHcURGQkptSHpIS0VGdz09
Denoising Diffusion Models: The Generative Learning Champion of the 2020s
Arash Vahdat
Principal Research Scientist
NVIDIA
Abstract
For a long time, the generative learning field especially around image generation was divided into two schools of thought: (1) generative adversarial networks (GANs) that generate high-quality samples at the cost of poor mode coverage and unstable training, and (2) likelihood-based models including variational autoencoders (VAEs), normalizing flows, and autoregressive models that provide full mode coverage often at the cost of poor sample quality. Recently, a new player called the denoising diffusion model has entered the generative learning arena, addressing the struggles of the two schools with both high sample quality and full mode coverage. However, these also come at the cost of slow sample generation.
In this talk, I will briefly review denoising diffusion models and some of the successful frameworks that we have recently built at NVIDIA using these models ranging from text-to-image generative models to 3D shape models and adversarially robust classification frameworks. I will then dive deep into the sampling challenges from diffusion models and discuss three frameworks we have developed for addressing them. These include latent score-based generative models that train diffusion models in a latent space, denoising diffusion GANs that use complex multimodal distributions for denoising, and higher-order solvers that solve the sampling differential equations in diffusion models in fewer steps.
Papers
https://arxiv.org/abs/2106.05931 (project page)
https://arxiv.org/abs/2112.07804 (project page)
https://arxiv.org/abs/2210.05475 (project page)
Biography
Arash Vahdat is a principal research scientist at NVIDIA research specializing in machine learning and computer vision. Before joining NVIDIA, he was a research scientist at D-Wave Systems where he worked on deep generative learning and weakly supervised learning. Prior to D-Wave, Arash was a research faculty member at Simon Fraser University (SFU), where he led research on deep learning-based video analysis, and taught master courses on machine learning for big data. Arash's current areas of research include deep generative learning, representation learning, and efficient sample generation.
|
Event Information
