Clued-in to Clueless: Navigating Distribution Shifts with Varying Levels of Target Distribution Information

Abstract: Distribution shifts, where deployment conditions differ from the training environment, are pervasive in real-world AI applications and often undermine model performance. This talk explores why distribution shifts present such challenges and offers actionable strategies to mitigate their impact. I will introduce modern principles and practices for navigating distribution shifts with varying levels of information about the target distribution, including unlabeled samples and shared causal structure. Highlights will include state-of-the-art approaches, from proxy methods for domain adaptation to in-context learning with foundation models. Additionally, I will examine the interaction between distribution shifts and AI policy, focusing on governance practices that enhance AI reliability in dynamic environments.

Speaker Bio: Olawale (Wale) Salaudeen is an incoming postdoctoral associate at MIT CSAIL starting in the fall of 2024. He completed his PhD in Computer Science at the University of Illinois at Urbana-Champaign, where he was affiliated with the Stanford Trustworthy AI Research (STAIR) Lab. Wale’s research focuses on developing machine learning methods for reliable and trustworthy real-world decision-making, specifically robust generalization, domain adaptation, and evaluation under distribution shifts. His work spans diverse applications, including neuroscience/neuroimaging, healthcare, and algorithmic fairness.

His work has been supported by a Sloan Scholarship, Beckman Institute Graduate Research Fellowship, GEM Associate Fellowship, and an NSF Miniature Brain Machinery Traineeship. During his PhD, he interned at Sandia National Laboratories, Google Brain, Cruise LLC, and the Max Planck Institute for Intelligent Systems.