Contextualized embeddings, and models built with them, command state-of-the-art performance across NLP. But such models can be opaque monoliths whose behavior can be hard to understand or control.
In this talk, first, I will present recent work on two different strategies to dissect the performance of such models. The first strategy takes a “behavioral” approach by crafting controlled input variations that should provoke specific model behaviors. Model outputs for such inputs can reveal reasoning failures and point to invalid heuristics that the models may have learned. The second dissection strategy uses geometric properties of embeddings to examine how concepts are organized in the high dimensional vector spaces. Doing so can tell us how these embeddings support various NLP tasks, and shed light into what
happens to them during fine-tuning.
Then, I will discuss the question of how we can guide these models to satisfy specific properties that are expressed in logic. To this end, I will show how we can systematically compile declarative rules into regularizers that guide their learning. I will present experiments which show that adding such declarative rules gives models are not only more accurate, but also more self-consistent in their predictions.
Vivek Srikumar is an associate professor in the School of Computing at the University of Utah and a visiting researcher at the Allen Institute for AI. His research lies in the areas of natural language processing and machine learning, and has been primarily driven by questions arising from the need to efficiently reason about textual data with limited supervision. His research has been published at various AI, NLP and ML venues, and has been recognized by a paper award at EMNLP 2014, and honorable mentions from CoNLL 2019 and the IEEE Micro magazine. His work has been supported by research grants from NSF, US-Israel BSF, NIH, and awards from Google, Intel, Nvidia and Verisk. He has organized several workshops hosted at the primary ML and NLP conferences around the theme of how learning and structured knowledge intersect. Furthermore, he has served as associate program chair of AAAI 2022 and the program co-chair of CoNLL 2022. He was a post-doctoral scholar at Stanford University before moving to Utah, and prior to that, in 2013, he obtained his PhD from the University of Illinois at Urbana-Champaign.
YOU ONLY NEED TO REGISTER ONCE TO ATTEND THE ENTIRE SERIES – We will send you email announcements with details of the upcoming speakers.
Register in advance for this webinar: https://usc.zoom.us/webinar/register/WN__0VhakI6Q6i3JsasdmNWcA
After registering, you will receive an email confirmation containing information about joining the Zoom webinar.
IF speaker approves to be recorded for this AI Seminar talk, it will be posted on our USC/ISI YouTube page within 1-2 business days: https://www.youtube.com/user/USCISI.
Host: Muhao Chen, POC: Maura Covaci