Efficient and effective deep learning in "supervision-denied" multisensory environments
Abstract
Deep learning has made dramatic advances in basic perceptual abilities, most notably in the realms of speech recognition and visual"" understanding. For domains where a large amount of explicit and accurate supervision is available, performance has been quite remar""kable, and consistently continues to increase. However, such methods~when conventionally trained~require on the order of millions"" of training examples to learn a network model. While this has been feasible in certain settings with a specific set of sensors, e.g""., RGB imagery in ImageNet, it is difficult at best and impossible in general for arbitrary novel sensor suites. Many such domains a"re therefore termed ~supervision-denied~. We propose to develop novel approaches based on our recently proposed Modality Transfer paradigm to provide strong learning in supervision-denied settings. We will provide effective scaffolding for deep learning models to" operate in domains with limited or no in-domain training data, unusual sensor types, or even dynamicallychanging sensor types or p""erformance characteristics. Our target application scenarios include pixel-level semantic segmentation of outdoor scenes into e.g.,"" road, vegetation, and object classes to support on and off road autonomous driving scenarios.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2017
- Source ID
- N000141712502
Entities
People
- Trevor Darrell
Organizations
- Office of Naval Research
- United States Navy
- University of California Regents