Towards resource efficient large neural networks
Bryan Plummer
IRB-3137
Abstract
To deploy AI agents in real world applications often requires considering number of factors that may not arise during relatively sanitized development stages like addressing domain shifts between training data and inference time, adapting to new categories, and performing efficient inference. Many see large models trained on web-scale datasets as a way to combat many of the issues with deploying models as their scale means that a larger portion of the target distribution may be seen that effectively reduces domain gaps while also reducing the potential for dataset bias during pretraining stages. In this talk, I will showcase some recent work in my lab where we demonstrate that many of these issues still remain despite the scaling effects, and new issues are introduced. In particular, we find that recent large model training simply shifts what data can be considered in-domain rather than making models more inherently generic. Relying on web-scraped datasets also produces models more vulnerable to attacks by web artifacts, intensifying the need for mitigating strategies. Finally, I will close by discussing methods to address energy, storage, and annotation issues stemming from adapting large models to a task.
Bio
Bryan Plummer is an Assistant Professor in the Department of Computer Science at Boston University and is a core faculty member of the Artificial Intelligence Research (AIR) Initiative in the Rafik B. Hariri Institute for Computing and Computational Science & Engineering. Bryan received his PhD from the University of Illinois at Urbana-Champaign in 2018 advised by Svetlana Lazebnik where he received a 3M Foundation Fellowship and was an NSF GRFP honorable mention. He was a Postdoc and a Research Assistant Professor at Boston University before taking his current tenure-track position in 2020. His research interests include multimodal reasoning, detecting manipulated and machine generated media, efficient neural networks, fair and explainable AI, and disentangled and structured representation learning. He has been an area chair for machine learning venues such as CVPR, ECCV, ICCV, NeurIPS, an action editor for ACL ARR, and a member of the editorial board of the International Journal of Computer Vision. He was named a 2023 Hariri Institute Junior Faculty Fellow and has been published in top-venues in machine learning, computer vision, and natural language processing.
This talk is organized by Samuel Malede Zewdu