Approximate-Analysis-by-Synthesis: Compositional Generative Networks with weak supervision, multi-task consistency and out-of-distribution performance.
Abstract
We propose a new class of generative models, which we call 3D Compositional Generative Networks (CGNs), which perform approximate-an alysis-by-synthesis whose performance is as good as Deep Nets on current benchmarks using standard performance measures (SPMs) but w hich can outperform them on more challenging tasks, such as out-of-distribution testing, multi-task consistently, and robustness to adversarial examiners, while requiring only weak supervision. We are particularly interested in enabling CGNs to detect parts, attri butes, and 3D structures of objects because these abilities will be of great practical use and, in particular for systems that combi ne vision with language and robotics. A key aspect of our research strategy, which exploits the ability of CGNs to perform modal-dom ain-transfer between real and synthetic images is the use of Computer Graphic models not only for representing objects but also to p rovide synthetic stimuli to help train and test them.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 07, 2021
- Source ID
- N000142112812
Entities
People
- Alan Yuille
Organizations
- Johns Hopkins University
- Office of Naval Research
- United States Navy