Talk: Learning under Resource Constraints in Real World

Title

Learning under Resource Constraints in Real World

 

Abstract

The current learning theory does not answer how to achieve the minimum statistical risk, i.e., optimal learning outcome under resource constraints and realistic settings. Meanwhile, the current set of practical tools to address resource constraints and realistic settings suffer from lack of generality and mainly focus on specific resource constraints and data-system perturbations. Constraints of computational platforms, data, and learning algorithms refer to limited resources in terms of computational resources, limited bandwidth that bounds communication costs among GPUs/nodes, the amount of available labeled data to learn, energy, and privacy budget. Realistic settings refer to potential statistical and system discrepancies as opposed to an ideal learning setting, e.g., distribution shifts, adversarial attacks,  hardware failures, and  slow machines. This talk presents conceptual results and working examples to minimize the statstical risk under constraints and realistic settings in ML.

 

Bio

Ali Ramezani-Kebrya is an Associate Professor in the Department of Informatics at the University of Oslo (UiO), a member of the ELLIS Society, a PI at the SFI Visual Intelligence, and the Norwegian Center for Knowledge-driven Machine Learning (Integreat). He has worked in ML at Vector Institute in Canada, EPFL in Switzerland, and Aalborg University in Denmark, co-authored in top ML venues including senior authorship in NeurIPS, received the Natural Sciences and Engineering Research Council of Canada Postdoc Fellowship. He is a Scientific Advisor for several EU companies and a Program Chair in the Northern Lights Deep Learning Conference 2024.