Bridging the Generalization and Interpretation Gaps in Deep Neural Networks

This project focuses on enhancing the reliability and understandability of advanced artificial intelligence (AI) systems, specifically deep neural networks (DNNs) - a type of AI that uses layered structures of interconnected elements, or "neurons," to process and interpret data in sophisticated ways. Currently, these AI systems face two main challenges. Firstly, they may struggle to apply what they've learned in one situation to a different one. This issue, known as the "generalization gap," is similar to a student who has crammed for an exam but struggles to apply the knowledge in a real-world scenario. Secondly, DNNs, like many AI systems, work in ways that can be difficult for humans to understand. This "interpretation gap" is like using a complicated machine without a user manual, which can make it hard to correct mistakes or explain why specific decisions were made. These challenges could have implications for any sector where AI is used, from healthcare to autopilot. If AI makes mistakes because of the generalization gap or if it's not clear why a decision was made due to the interpretation gap, it could lead to significant errors, lack of trust, or even potential harm. This project aims to study these issues, enhancing the reliability and transparency of AI systems, enabling us to apply these technologies more confidently and effectively. By doing so, it will advance our scientific understanding of AI, support education in this vital field, and benefit society by ensuring AI technologies are more dependable and understandable. The project also provides research opportunities for graduate students. This project aims to develop a new framework to address the generalization and interpretation gaps in DNNs by investigating a series of well-defined research problems. The work includes the development of novel statistical theories for a better understanding of generalization errors in DNNs, the creation of robust and computationally efficient algorithms, and the promotion of innovative approaches for out-of-distribution generalizations. This project will advance our understanding of DNNs, develop new methods and algorithms, and provide insights into practical applications in diverse fields. The principal investigators will incorporate the research findings of this project into their educational endeavors. 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.