Watch a conference talk from Harvard CMSA's Big Data Conference where MIT's Tommi Jaakkola explores generative modeling and physical processes, focusing on deep distributional modeling techniques for complex generative tasks in natural sciences and engineering. Learn about diffusion models and electrostatic models (Poisson flow), understanding their relationships in terms of embedding dimension. Discover recent developments in SE(3) invariant distributional modeling for backbone 3D structures, including the generation of designable monomers without pre-trained protein structure prediction methods, and state-of-the-art image generation capabilities through Poisson flow. Gain insights into the efficiency analysis of sample generation in these models, covering topics from introduction and overview to self-consistency evaluation, diffusion models, image generation, and synthetic evaluation.
Generative Modeling and Physical Processes - From Diffusion Models to Electrostatic Applications
Harvard CMSA via YouTube
Overview
Syllabus
Introduction
Overview
Where do these models come from
Generative tasks
Local frames
Recap
Reverse diffusion process
Self consistency evaluation
Pretraining
Diffusion Model
Image Generation
Synthetic Evaluation
Generative Models
Sample sampler
Summary
Taught by
Harvard CMSA
Related Courses
-
Protein Design with Generative Diffusion Models
-
Diffusion Based Distributional Modeling of Conformers, Blind Docking and Proteins
-
PFGM++: Unlocking the Potential of Physics-Inspired Generative Models
-
GenPhys: From Physical Processes to Generative Models
-
Reflected Diffusion Models: Principled Data Constraints in Generative Modeling
-
Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport
