Fundamentals and Applications in Learned Information Representation
Abstract
This project addresses fundamentals in information representation, developing both theory and state-of- the-art applications ranging from 3D-LiDAR classication to multi-scale object recognition with nested condence. We start by developing a framework to use basis functions as the building lters in deep neural networks. Such basis functions lead to natural invariant representations, domain transfer, and stochasticity in the representation and network; and are plug-and-play tools that can be used in virtually all architectures and tasks. Beyond the fundamentals and computational aspects, each part of the project includes numerous applications in the elds of image/video and natural language processing, and many are addressed as part of this project as well. Beyond the particular challenges and applications in this project, such learned invariant and robust representations provide a framework to connect with other unique aspects of machine learning, from causality to privacy. This is thereby a step towards providing a unied perspective for data science and machine learning. With the contributions in this project we bring machine learning in general and deep learning in particular to a level of understanding and performance needed for environments with limited data for training and requiring condent deployment.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 08, 2020
- Source ID
- N000142012339
Entities
People
- Guillermo Sapiro
Organizations
- Duke University
- Office of Naval Research
- United States Navy