Georgia Institute of Technology

Featuring Karik Goyal, Georgia Institute of Technology

Abstract: I will talk about the challenges and potential solutions for endowing probabilistic generative models with controllability. First, I will talk about language modeling and estimating conditional distributions under control attributes. Specifically, I will describe how despite the difficulty of working with these conditional distributions, we use them to uncover statistical properties related to search and sampling in language models. Next, I will talk about our work on custom controllable and interpretable generative models for modeling early modern (1600s) printed text to discover new historical knowledge around censorship and intellectual networks in this period.

Bio: Kartik Goyal is an assistant professor in the School of Interactive Computing at Georgia Tech. His research interests are in natural language processing and machine learning problems that involve developing probabilistic models of latent structure in naturally occurring data with capabilities of interpretability and control. His recent work has focused on developing machine learning techniques for various problems in cultural analytics and digital humanities. Prior to Georgia Tech, he was a research assistant professor at Toyota Technological Institute at Chicago. He received his PhD from Language Technologies Institute at Carnegie Mellon University.