IEEE Information Theory Society Distinguished Lecture: Understanding and using the embedding spaces of large generative models : vTools Events

IEEE.org | IEEE Xplore Digital Library | IEEE Standards | IEEE Spectrum | More Sites

IEEE Information Theory Society Distinguished Lecture: Understanding and using the embedding spaces of large generative models

#information-theory #information-theory-society #machine-learning #distinguished-lecture #embedding-spaces

Title: Understanding and using the embedding spaces of large generative models

Speaker: Anand Sarwate, Professor of ECE, Rutgers University

Abstract:

Training massive ML/AI models on massive amounts of data supposedly creates latent representations or features that are “universal” in the sense that the large model acts as a feature extractor that maps inputs into an embedding space. In this talk I will discuss recent empirical work that looks at embeddings for generative models. In particular, I will describe an approach that uses a third model as a “microscope” to uncover differences between two other models. Simple methods on the embedding space of the “microscope” model show that outputs of different models are distinguishable, which potentially opens the door to several applications. Time permitting, I will describe other insights about embedding spaces.

Bio:

Anand D. Sarwate is currently a Professor of Electrical and Computer Engineering at Rutgers University. Prior to joining Rutgers he was a Research Assistant Professor at TTI-Chicago and a postdoc at the ITA Center at UCSD. He received undergraduate degrees in mathematics and electrical engineering from MIT (2002) and a PhD from UC Berkeley (2008). His research interests include information theory, machine learning, signal processing, optimization, and privacy and security. Dr. Sarwate is a Distinguished Lecturer of the IEEE Information Theory Society for 2024--2025 and is on the Board of Governors of the IEEE Information Theory Society.

Location:

University of Illinois Chicago, Lecture Center C4