Simpler to build and more accurate diffusion models for generation of real world data
Abstract
Approved for Public Release.Progress in artificial intelligence (AI) pervades everyday life through improved medical diagnostics tobetter machine translation. Recently, a new style of AI has taken hold called generative AI. Generative AI is a collection of algorithms that create objects ranging from pictures, to molecules, to even computer code. This project seeks to advance the state of a special type of generative AI that, first, purposely corrupts a dataset, and second, learns to undo this process thereby creating newsimilar data. This class of generative AI is called diffusion models. While diffusion models have produced state-of-the-art resultsin generation of images, text, and audio, they are slow and expensive to train, require cumbersome mathematical derivations, and are restrictive in their construction. This project aims to develop methods for training diffusion models that are faster and simpler to use, and strengthen the foundations of understanding what makes a good generative AI technique.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 24, 2023
- Source ID
- N000142312634
Entities
People
- Rajesh Ranganath
Organizations
- New York University
- Office of Naval Research
- United States Navy