Towards More Reliable Generative AI: Probing Failure Modes, Harnessing Test-Time Inference, and Interpreting Diffusion Models

Abstract:
In this talk, we explore strategies for making generative AI more trustworthy by identifying its vulnerabilities, improving interpretability, and employing adaptive techniques. First, we introduce MediConfusion, a new benchmark that uncovers systemic failure modes in state-of-the-art medical multimodal models for Visual Question Answering—a domain where accuracy is critical. We then address interpretability challenges in diffusion models, shedding light on how and why these powerful generative approaches produce their outputs. Finally, we discuss test-time training (TTT), a gradient-based method that adapts model parameters for each test instance, offering both a theoretical and empirical framework that demonstrates how TTT can significantly reduce sample requirements for in-context learning. By uniting insights from model failure analysis, interpretability in diffusion-based generative models, and test-time adaptation, we chart a path toward more reliable generative AI systems for healthcare and beyond.

Bio:
Mahdi Soltanolkotabi is the director of the center on AI Foundations for the Sciences (AIF4S) at the University of Southern California. He is also a professor in the Departments of Electrical and Computer Engineering, Computer Science, and Industrial and Systems engineering. Prior to joining USC, he completed his PhD in electrical engineering at Stanford in 2014. He was a postdoctoral researcher in the EECS department at UC Berkeley during the 2014-2015 academic year.
Mahdi is the recipient of the Information Theory Society Best Paper Award, Packard Fellowship in Science and Engineering, an NIH Director’s new innovator award, a Sloan Research Fellowship, an NSF Career award, an Airforce Office of Research Young Investigator award (AFOSR-YIP), the Viterbi school of engineering junior faculty research award, and faculty awards from Google and Amazon. His research focuses on developing the mathematical foundations of modern data science via characterizing the behavior and pitfalls of contemporary nonconvex learning and optimization algorithms with applications in AI, deep learning, large scale distributed training, federated learning, computational imaging, and AI for scientific and medical applications. Most recently his applied research is focused on developing and deploying reliable and trustworthy AI in healthcare.