CAREER: Continual Learning with Evolving Memory, Soft Supervision, and Cross-Domain Knowledge - Foundational Theory and Advanced Algorithms

As the landscape of data and learning environments expands rapidly, the utility of artificial intelligence hinges on its ability to scale and adapt. In response, continual learning is emerging as a promising paradigm to meet this demand. Distinct from other learning paradigms, continual learning emphasizes the ability to maintain performance on previously learned tasks while seamlessly integrating new information. It focuses on rapid adaptation to new environments by appropriately recalling past knowledge and actively seeking side information to accelerate learning and improve accuracy. However, current methods in continual learning are predominantly empirical and lack a clear theoretical foundation, limiting their wider application and hindering further progress. This project aims to bridge this gap by developing a principled framework for continual learning, consisting of novel theoretical insights and practical algorithmic designs. The research outcomes will substantially advance our understanding of continual learning, enrich its algorithmic framework, and provide scalable algorithm packages. The outcomes will be integrated into machine learning courses at all levels, benefiting students across disciplines such as electrical engineering, statistics, and computer science. The project will actively involve underrepresented students in STEM, synergizing research and education at undergraduate and graduate levels, and developing introductory materials for K-12 students through AI apprenticeship programs. The overarching goal of this project is to pioneer a methodological framework of continual learning for autonomous machine learners navigating dynamic data environments. The framework addresses three essential facets of continual learning. First, it will establish mathematical foundations of continual learning and scalable online algorithms to recollect an evolving memory for fast adaptation and capability expansion. Second, it will develop methods to amplify learning efficiency from informative, even less reliable, pseudo labels. Lastly, it will develop approaches for expanding the predictive power with task-specific assistance from peer learners who possess cross-modal data. The developed foundations will lead to novel algorithms to empower intelligent engineering systems to learn more efficiently and effectively in an ever-changing world. The project will also offer deep insights into pivotal machine learning questions, such as understanding the fundamental trade-offs between adaptivity and forgetting in dynamic learning environments, gauging the maximal influences of minor yet jointly dependent data perturbations on modeling results, and quantifying a model's room for improvement when the underlying law of probability remains elusive. Furthermore, these developments are anticipated to have broader impacts, enabling resource-constrained machine learners to continually broaden their learning capabilities in complex practical applications, such as intelligent transportation systems and cellular performance management. 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.