Visual Learning and Reasoning from Incomplete Information

Abstract

To improve decision-making in complex real-world environments, we will advance the state-of-the-art in learning representations of objects and activities from visual data, as well as forecasting uncertain future actions and intentions. First, to ease deployment of computer vision methods in real-world naval scenarios, we will develop novel algorithms for weakly supervised learning. These methods will improve learning in cases where many images or videos from the target environment are available, but few of them have thetarget labels (e.g., semantic object or activity categories) of interest. They will also improve the ability of models trained on synthetic data, as generated by modern computer graphics rendering methods, to generalize to real imagery. Second, to enable temporal forecasts that account for uncertainty and incorporate domain knowledge, we will employ principled hybrids of deep neural networks and probabilistic graphical models. These stochastic dynamical models will enable our novel particle-based framework for reasoning about multiple possible future trajectories for entities with complex behaviors.It is understood that any developmental items and specially designed parts, components, accessories and attachments generated under this Defense Department agreement are being developed for both civil and military applications.

Document Details

Document Type
DoD Grant Award
Publication Date
Jul 24, 2023
Source ID
N000142312712

Entities

People

  • Erik Sudderth

Organizations

  • Office of Naval Research
  • United States Navy
  • University of California, Irvine

Tags

Fields of Study

  • Computer science

Readers

  • Computer Vision.
  • Distributed Systems and Data Platform Development
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - DoD AI Strategy
  • AI & ML - Neural Networks