The Soc(AI)ety Seminars, hosted by the Lucy Family Institute for Data & Society, are a collection of talks with a vision for AI’s present and future impact on society. Each session is meant to inspire a dialogue on ethical and socially responsible Data & AI innovation. For more information, and to view previous Soc(AI)ety Seminars sessions, please visit the Soc(AI)ety Seminars webpage.
Description:
Generative models have shown impressive capabilities in synthesizing high-quality outputs across various domains. However, a persistent challenge is the occurrence of “hallucinations,” where the model produces outputs that are not grounded in the underlying facts. While empirical strategies have been explored to mitigate this issue, a rigorous theoretical understanding remains elusive. In this talk, I present a theoretical framework to analyze the learnability of non-hallucinating generative models from a learning-theoretic perspective. Our results reveal that non-hallucinating learning is statistically impossible when relying solely on the training dataset, even for a hypothesis class of size two and when the entire training set is truthful. To overcome these limitations, we show that incorporating inductive biases aligned with the actual facts into the learning process is essential. We provide a systematic approach to achieve this by restricting the fact set to a concept class of finite VC-dimension and demonstrate its effectiveness under various learning paradigms. Although our findings are primarily conceptual, they represent a first step towards a principled approach to addressing hallucinations in learning generative models.
This work is in collaboration with Profs. Changlong Wu and Wojciech Szpankowski.
Guest Speaker Bio:
Ananth Grama is the Samuel Conte Distinguished Professor of Computer Science at Purdue University, Director of the Institute for Physical Artificial Intelligence, and Associate Director of the Center for Science of Information, a Science and Technology Center of the National Science Foundation. Grama received his Ph.D. in Computer Science in 1996 from the University of Minnesota and has been at Purdue since.
Grama’s research interests include parallel and distributed computing, AI models and methods, and applications. Ananth’s work has been recognized though a number of awards, including the Purdue University Outstanding Assistant Professor (1998), the NSF CAREER Award (1998), Purdue University School of Science Outstanding Teacher Award (2002), Purdue University Faculty Scholar (2002), Most Influential Professor in Computer Science (2010), Fellow of the American Association for the Advancement of Science (AAAS) (2013), Distinguished Alumnus Award from the University of Minnesota (2015), and Amazon Research Award (2021).