CAREER: AdaptTrust: Adaptable, Trustworthy and Uncertainty-Aware Learning in Intelligent Sensing Technologies

This NSF CAREER project aims to develop machine learning systems that continuously learn from new data without forgetting prior knowledge. Today’s machine learning models are powerful but largely static, often overwriting earlier information when updated, a problem known as catastrophic forgetting that reduces reliability in evolving environments, such as healthcare monitoring, environmental sensing, and autonomous systems. The project will bring transformative change by enabling intelligent systems to recognize what they do not know, quantify uncertainty, update safely to new information, and retain prior knowledge without becoming unstable or overconfident. This will be achieved by designing learning methods that leverage uncertainty as a guiding signal to manage when models should preserve knowledge, adapt, or expand internal capacity. The intellectual merit includes advancing fundamental theory and algorithms for continual learning, uncertainty-aware decision-making, and modular neural architectures to enable stable, robust, and scalable lifelong learning. The broader impacts include advancing trustworthy AI that supports national priorities, including healthcare innovation, environmental monitoring, and intelligent infrastructure; strengthening economic competitiveness through adaptive intelligent technologies; supporting resilient sensing relevant to national defense; and integrating research with education from K-12 through graduate levels to promote scientific progress and prepare a skilled STEM workforce. This project develops a unified, theoretically grounded framework for continual learning from evolving data distributions without storing past data. The proposed research combines Bayesian uncertainty quantification, adaptive optimization, and modular neural architectures to enable continual learning in dynamic environments. Key objectives include: (1) uncertainty-guided optimization methods that regulate parameter updates across sequential tasks with theoretical generalization guarantees; (2) neural self-management masking strategies that selectively preserve or adapt parameters; (3) modular mixture-of-experts architectures with uncertainty-aware gating for scalable routing, and capacity expansion; and (4) convex-relaxed optimization formulations favoring flat minima under noise and domain shifts. The framework will be validated on benchmark and real-world sensing datasets. The integrated research-education program translates scientific advances into structured training across educational levels. The project will provide individualized mentoring and research opportunities for graduate, undergraduate, and high-school students; develop research-integrated curricula spanning foundational mathematics, robust machine learning, and adaptive sensing; and implement hands-on projects using real-world data through industry collaborators. Workforce development will be supported through workshops and tutorials, and K–12 engagement will offer experiential learning through guided projects on trustworthy and adaptive AI for environmental monitoring. These activities will help prepare the next generation of scientists and engineers to develop reliable, adaptive intelligent systems that serve societal needs. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2542166 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Dimah Dera | Institution: Rochester Institute of Tech, ROCHESTER, NY | Award Amount: $560,000 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2542166 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2542166.html