Adaptive, Robust and Informed Algorithms for modern Machine Learning (ARIA-ML)
Abstract
Machine learning is a key enabling technology for modern data analysis. ARIA tackles key challenges and limitations of state of the art to develop a new generation of provably reliable and efficient machine learning methods. State-of-the-art AI systems indeed provide impressive results, opening new horizons for science, technology, and society, however a close look at the resources needed is worrisome. AI systems rely on a massive amount of data, either manually annotated or generated by expensive simulations. Further, they often rely on the brute force empirical exploration of multiple stacks of computational layers. While performances are good they are brittle. This AI growth model poses challenges that endanger the potential benefits. Understanding how to make AI sustainable, and tackle the above challenges, requires rethinking the way AI systems are developed by designing, studying and deploying a new generation of AI algorithms that are: adaptive to the nature of the data and the model used for learning, robust with respect to both stochastic/adversarial perturbations of the data and model misspecifications, and informed by the prior information available in terms of structural and evolution equations. In this project we will work to implement these objectives developing solutions that are at the same time theoretically sound and providing good empirical performances.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 20, 2023
- Source ID
- FA86552217034
Entities
People
- Silvia Villa
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Genoa