CAREER: PTM-SEER: Software Engineering Foundations for Re-Using Pre-Trained Neural Models

Modern computing systems increasingly incorporate learned components using techniques from machine learning and artificial intelligence. Engineering practice favors reuse over building from scratch. However, while for conventional software we know much about the re-use and adaptation of components, the correspondence for pre-trained models is an emerging and evolving concern. Engineers must decide which models to trust, how to adapt them, and how to document their behavior, often without shared standards or guidance. The project’s novelties are its systematic investigation of how model reuse resembles, and differs from, traditional software reuse, and the creation of practical methods that make these differences manageable. The project's broader significance and importance are reflected in a toolkit that enables more efficient engineering practices, lowering the costs of developing intelligent computing systems. The project also produces substantial educational materials to support K-12, undergraduate, and graduate students, as well as practicing professionals. Its result is improved United States economic competitiveness, greater academia-industry partnerships, and a deeper pipeline of engineers with AI skills for opportunities in industry, academia, and government. The project applies methods from human factors and software systems engineering to study how practitioners discover, evaluate, adapt, and maintain pre-trained models. It identifies best practices, constructs taxonomies of engineering behaviors, and develops novel tools to accelerate software engineering work. The resulting knowledge covers the full re-use cycle, including (1) techniques to facilitate the identification of pre-trained models; (2) techniques to support the evaluation and selection of such models; (3) a novel, ecosystem-spanning dataset of models for further analysis; and (4) a grounded-theoretic advance on software engineering theory that contrasts the reuse of statistical, data-dependent learned models with conventional software --- altering assumptions about modularity, specification, and verification. Together, the project provides a foundation for principled, efficient, and trustworthy reuse of artificial intelligence components in modern computing systems. 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: 2541917 | Program: 01003031DB NSF RESEARCH & RELATED ACTIVIT,01002930DB NSF RESEARCH & RELATED ACTIVIT,01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: James Davis | Institution: Purdue University, WEST LAFAYETTE, IN | Award Amount: $401,546 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2541917 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2541917.html