Fragility and trust in autonomous systems
Abstract
While neural networks have achieved state of the art performance under normal conditions, they exhibit disqualifying fragility under some less controlled environments. This white paper proposes several research directions intended to fortify properties of neural networks that currently prevent them from being used in real-world scenarios. We propose impactful research intended to allow neuralnetworks to handle common mismatches between training and deployment datasets, to avoid dangerous output that humans know to avoid while the networks do not, to recognize biased and nondiverse datasets, and be less vulnerable to effective data poisoning attacks.Approved for Public Release
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 09, 2021
- Source ID
- N000142112557
Entities
People
- Tom Goldstein
Organizations
- Office of Naval Research
- United States Navy
- University of Maryland