Ambiguity-aware Artificial Intelligence via Statistical Inference
Abstract
Recent successes of machine learning (ML) and artificial intelligence (AI) in computer vision and natural language processing have garnered a great deal of attention. At the same time, ML and AI systems are built for straightforward prediction tasks by design and are not designed to handle ambiguity and uncertainty. There is a lot of empirical evidence that such methods can fail in ambiguous situations. In this project, we aim to develop methods that can help machine learning systems better deal with ambiguity, by leveraging ideas from statistical inference. Statistical reasoning has been successful for more than a hundred years in helping deal with noisy, ambiguous data. We aim to leverage such ideas (e.g., confidence intervals, prediction sets) and adapt them to modern machine learning and artificial intelligence scenarios; by also leveraging other notions and results from applied mathematics, including combinatorics and group theory. Our methods will be generally applicable to a broad range of learning problems and models and are expected to improve the reliability of machine learning and artificial intelligence systems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 06, 2025
- Source ID
- FA95502410150
Entities
People
- Edgar Dobriban
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Pennsylvania