Probabilistic Label-Efficient Deep Generative Structures (PLEDGES)
Abstract
The enormous commercial success of machine learning has failed to translate into high performance military applications. Though deep learning is beginning to show impressive results in several specific military tasks, current capabilities fail to perform sufficiently due to their requirement for extremely large, labeled training sets. Under the Probabilistic Label-Efficient Deep Generative Structures (PLEDGES) project, part of DARPA's Learning with Less Labels (LwLL) program, we pursued pioneering research at the interface of probabilistic modeling and deep learning. We investigated three main directions. First, we developed Structured Deep Probabilistic Models (SDPMs), which define structured and disentangled joint probability distributions over unlabeled data observations. Second, we pursued efficient and accurate algorithms for high-capacity probabilistic models. Third, we developed probabilistic variants of deep learning models for semi-supervised and weakly supervised learning.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 11, 2021
- Accession Number
- AD1145098
Entities
People
- Avi Pfeffer
- Brad R Rosenberg
- Catherine Call
- Deniz Erdoğmuş
- Frank Wood
- Ishaan Shah
- Jan W. Van De Meent
- Kirstin Bibbiani
- Leonid Sigal
- Sameer Singh
Organizations
- Charles River Analytics (United States)