Variational Mutual Information Estimation: from Data Collection to Large Vision-Language Models

Presented by Qing Guo, Virginia Tech

Tuesday, January 30 2024
3:30 PM-4:30 PM ET
AUST 105
Webex Meeting Link
Coffee will be served at 3:00 pm in the Noether Lounge (AUST 326)

In today's digital landscape, the deluge of data across various domains can easily overwhelm traditional learning methods to extract useful information. Central to these challenges is the efficient quantification of general associations between variable pairs, and the accurate utilization of this information in decision-making processes. In my talk, I will delve into the critical role of Mutual Information (MI) in modern data science and discuss scaling MI estimation with deep neural networks to address the complexity of contemporary datasets. Specifically, I will present our recent work proposing a novel MI estimation framework powered by variational inference, contrastive estimation, and convex optimization. The proposed method provides theoretical guarantees on the convergence of variational non-parametric MI estimates and features simple implementations. The effectiveness of the proposed method will be elaborated in several applications, including Bayesian optimal data collection, self-supervised learning, conversational recommendation systems, and large vision language models.

Speaker Bio:

Qing Guo is from Department of Statistics at Virginia Tech (VT). She is also a member of VT Statistics and Artificial Intelligence Laboratory (VT-SAIL). In 2022, she was honored as an Amazon Fellow. Her research focuses on addressing some of the fundamental challenges in machine learning using novel mathematical and statistical insights, with both theoretical analysis and efficient algorithms. Her current focus revolves around enhancing the data efficiency and robustness of artificial intelligence systems through the integration of ideas from information theory. Her research encompasses various topics that comprehensively span the life cycle of AI models, including Data collection, Generative AI, Self-supervised learning, and Knowledge transfer.