AI Workshop Brings Researchers Together to Discuss Rapid Progression of Large Pre-Trained Models

The workshop helped to unite researchers from multiple disciplines to explore the future of technology research and foster collaboration.
Descriptive image for AI Workshop Brings Researchers Together to Discuss Rapid Progression of Large Pre-Trained Models

General-purpose generative models are the massive, big-data-driven systems powering new and exciting artificial intelligence technology. These systems use large language models, complex algorithms and neural networks to produce original text, audio, synthetic data, images and more.

However, the impressive performance of these models comes at a cost. They require significant data, computational power, and storage, creating a barrier to entry, especially for smaller research groups. The hype around large pre-trained models (LPMs)—a deep learning model that is trained on large datasets to accomplish a specific task—has also created a sense of panic in the research community about how they can stay up to date and contribute to progress.

A University of Maryland workshop, held on November 3 at the Brendan Iribe Center for Computer Science and Engineering, brought together researchers from across disciplines to discuss these challenges and the future of natural language processing, computer vision, human-computer interaction, and robotics research; as well as provide a forum to exchange ideas, share experiences, and form collaborations.

Navita Goyal, a third-year computer science doctoral student and lead organizer of “Riding the LPM Wave: The Future of Academic Research in the Era of Pre-Trained Models,” says the workshop provided a critical platform to address the challenges and ethical considerations associated with LPMs.

“In the ever-shifting landscape of generative models, academia’s unique role is not only in solving problems, but in defining the right questions, shaping the path to real-world applications, compliance, understanding, and revealing their potential for scientific discovery across domains,” says Goyal, who is also a member of the Computational Linguistics and Information Processing Lab (CLIP).

Hal Daumé III, a professor of computer science and member of CLIP with appointments in the Language Science Center and the University of Maryland Institute for Advanced Computer Studies (UMIACS), was the workshop’s faculty organizer.

“We organized this workshop—really the students organized this workshop—as a response to the quick pace of progress in AI in industry, and a sense of impending dread around what it means to be an academic doing AI these days," he says.

More than 70 UMD faculty members, postdocs and students attended the event, which was sponsored by UMIACS and the university’s Values–Centered AI Initiative (VCAI). Established with $700,000 in seed funding from the university’s Grand Challenges Grants Program, VCAI is a new multidisciplinary center that aims to establish UMD as a leader in the field of human-centered AI by producing high-quality research, developing an innovative curriculum, and serving as a resource for faculty and students across campus.

“I came out of the workshop more convinced than ever that academia—and UMD in particular—is a wonderful place to be doing AI, and that we can have a substantial impact while also benefiting from industry advances,” says Daumé, who is also the lead principal investigator for VCAI.

Jacob Andreas, an associate professor of electrical engineering and computer science at the Massachusetts Institute of Technology, was the workshop’s keynote speaker. He presented his recent research on two very different scientific problems arising from applications of language models (LMs): doing science with LMs—to understand the structure of sperm whale communication; and applying science on LMs— specifically to discover new algorithms for arithmetic and language learning.

A panel on LPMs consisted of Andreas, Daumé and Yizheng Chen, an assistant professor of computer science with an appointment in UMIACS who is also a core member in the Maryland Cybersecurity Center.

Chen says that the panel highlighted UMD researchers’ innovative thinking on a hot research topic, and that the event was a helpful resource for students’ professional development and future job prospects in the field.

In addition to Goyal and Daumé, the event’s organizers included sixth-year computer science doctoral student Yang (Trista) Cao; Zijian (Jason) Ding, a third-year information studies doctoral student; and John Kirchenbauer, a third-year computer science doctoral student.

—Story by Melissa Brachfeld, UMIACS communications group

The Department welcomes comments, suggestions and corrections.  Send email to editor [-at-] cs [dot] umd [dot] edu.