Generalization Under Distribution Shifts: From Algorithms to Applications

As machine learning systems are increasingly deployed in real-world environments, it becomes essential to ensure their generalizability and robustness under distribution shifts. These shifts arise when the test data differs from the training data in subtle or significant ways, often due to changes in context or measurement conditions. In this talk, I will present my research on designing theoretically grounded algorithms that enable models to generalize across diverse data distributions with robust performance. I will also discuss algorithms for achieving reliable generalization in more challenging and realistic scenarios, such as learning from noisy labels and under limited computational resources. Beyond algorithmic contributions, I will highlight the interdisciplinary applications of my work across scientific domains, including neuroscience, bioinformatics, and particle physics. I will conclude with my vision for building more scalable and trustworthy AI toward the long-term goal of advancing Artificial General Intelligence (AGI) that serves the broader good of society.

Song Wang, a fifth-year PhD candidate in ECE at the University of Virginia,advised by Professor Jundong Li, received a bachelor's degree in EE from Tsinghua University in 2020. His research focuses on the generalizability and robustness of machine learning algorithms under distribution shifts, enabling reliable performance across data distributions and real-world contexts. He also explores interdisciplinary applications of his work in scientific domains such as bioinformatics and neuroscience. His work has been featured as oral and spotlight presentations at top-tier conferences such as NeurIPS, ICLR, ICML, AAAI, and SIGKDD, and he is a recipient of the PAKDD 2024 Best Paper Award.