Learning Zero-Shot with Limited Supervision Data
Abstract
Conventional methods for training high-accuracy predictors and classifiers are based on supervised learning. Fully labeled training data is first acquired for each class. Then classifiers are trained on training data to classify new examples belonging to one of the observed classes seen during training at run-time. In many DOD scenarios the amount of annotated data is limited and sparse or unavailable for many target classes. Data collects spanning the wide-variety of target classes with associated human annotated ground truth is in general expensive. Furthermore, in many ISR scenarios, we often encounter new and evolving targetclasses, and for these cases it is not possible to acquire fully labeled training data.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 10, 2018
- Source ID
- N000141812257
Entities
People
- Venkatesh Saligrama
Organizations
- Boston University
- Office of Naval Research
- United States Navy