CAREER: Toward Interpretable AI: Modeling and Adapting Data-Driven Semantic Prototypes

Artificial intelligence (AI) is increasingly used to support decisions in areas such as healthcare, transportation, and environmental monitoring, yet many systems operate as black boxes, making it difficult for people to understand how decisions are made or to assess their reliability. This lack of transparency can lead to errors, reduce user confidence, and limit the safe and effective use of AI in situations where clear reasoning is essential. This project addresses these challenges by developing new AI methods that explain the decisions in straightforward, intuitive, and meaningful ways, enabling systems to identify and use recognizable patterns in data as building blocks for reasoning. By making AI more transparent and interpretable, the project will help professionals such as doctors and engineers make better-informed decisions, enhance safety in high-stakes applications, and strengthen public trust in AI technologies. In addition, the project includes education and outreach activities that train students in responsible AI development, helping to build a skilled workforce, broaden participation in technology innovation, and contribute to long-term economic growth and societal well-being. This project develops a novel framework for interpretable artificial intelligence (AI) that learns semantically meaningful data patterns and organizes them into transparent, compositional reasoning processes. The objective is to improve model interpretability by explicitly revealing how and why decisions are made, enabling systematic identification of errors, assessment of data quality, and development of more reliable models. The project focuses on designing architectures that capture representative, human-interpretable patterns while mitigating spurious correlations, and that generalize effectively to new tasks and evolving environments. The proposed framework will be integrated with modern deep learning paradigms, including convolutional neural networks (CNNs) and Vision Transformers, to ensure compatibility with state-of-the-art practice. It will be rigorously evaluated against established prototype-based interpretable methods using quantitative metrics of predictive accuracy, interpretability, robustness, and computational efficiency. Experimental validation will be conducted on standard benchmark datasets as well as real-world applications, including medical decision support and autonomous systems. The project also incorporates education and outreach by developing curriculum modules, open-source software, and interdisciplinary training opportunities, thereby advancing broader understanding and adoption of transparent and trustworthy AI 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: 2543694 | Program: 01002930DB NSF RESEARCH & RELATED ACTIVIT,01002627DB NSF RESEARCH & RELATED ACTIVIT,01003031DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Zhengming Ding | Institution: Tulane University, NEW ORLEANS, LA | Award Amount: $418,410 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2543694 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2543694.html