Professor of computer science at UdeM wins the most prestigious award in the United States

The US Department of Energy’s Office of Science has released the program’s list of 56 laureates. Incentive 2023, featuring a computer science professor from the University of Montreal Irina Rech.

The only Canadian in the 2023 vintage of the competition, m^me Rish was given 990,000 “knot hours” for the next 12 months in Supercomputer Summit At Oak Ridge National Laboratory, Tennessee. A “node hour” is the use of a node (or unit of computation) on a supercomputer for one hour.

M^me Reich will thus have the opportunity to continue his work on scalable core models for transferable general-purpose artificial intelligence (AI), aimed at training large multimodal models of deep neural networks.

“With this support, we not only have access to a supercomputing resource, but we are taking a step towards achieving our long-term goal, which is to democratize AI and prove that universities and the free software community can be competitive in this sector.”

Irina Resch is Professor in the Department of Computer Science and Operations Research at UdeM and a member of Mila, the Institute for Artificial Intelligence in Quebec. She holds the Canada Chair of Research Excellence in Independent Artificial Intelligence and the Canada-CIFAR AI Chair.

Twenty years at IBM

Irina Resch’s research focuses on artificial intelligence, machine learning, and neural data analysis. Professor at UdeM since 2019, she taught at Columbia University and worked for 20 years at IBM’s TJ Watson Research Center, where she collaborated on various projects in neuroscience and artificial intelligence.

The INCITE (Innovative and New Computational Impact on Theory and Experiment) program was established in 2003 by Raymond Lee Auerbach, then Under Secretary of State for Science, to help the research community advance science and engineering disciplines.

Applications for the program competition are evaluated according to a highly selective process that takes place over a period of four months. Expert and review panels scrutinize the projects from different angles in order to choose which teams will have access to one of the four Oak Ridge supercomputers for a year.

Of the 97 candidates evaluated this year, he was Half was awarded Time using the supercomputer, including Irina Resch and her collaborators Stella Biedermana master’s student at the Georgia Institute of Technology and a senior scientist in the EleutherAI Collective Research Group, W Genia Zaitseva senior researcher in the Multipurpose AI and Scalable Learning Lab at the Juelich Supercomputing Center, Germany, and chief scientist and founder of LAION, a non-profit organization focused on free AI.

Other contributors to the INCITE proposal are Quentin Anthony, Guillermo Cecchi, Mehdi Cherti, Guillaume Dumas, Eric Hallahan, Yonggang Ho, Sergey Panetkin, Christophe Schumann and Ryo Yokota.

As part of their work in Tennessee, in addition to training large neural network models, Irina Resch and her team will take advantage of this access to study scaling laws and emerging behaviors, and attempt to apply their general models and the knowledge gained to them in many practical processes. uses.

“In line with human values”

The team aims to develop AI towards broader interests while ensuring that it is risk-free and “in line with human values ​​- but also contributes to advances in other areas such as health and the biomedical sciences,” explains Irina Resch.

To do this, you need to develop large, powerful generic models that are already trained in an autonomous fashion from a variety of data sets. These models can serve as a transferable knowledge base and have much broader uses than AI currently allows.

Building on its recent successes in this area, the team intends to build large neural network models called “transformers” that have recently achieved impressive results in language modeling and image processing.

Irina Resch and her colleagues will evaluate the possibilities of overcrowding with larger models and already trained datasets. The team plans to extend these models to a large number of applications beyond text and images as well as various machine learning tasks.

“The goal is to develop them into adaptive systems that are constantly learning and perfecting to derive predictive models that can be applied to different uses such as caregiving and brain imaging,” says Irina Resch. And all of this, we are going to make it publicly available for everyone to enjoy. Science is for the masses, and this project is a perfect example of that.”