Colloquium: Distributed Asynchronous Stochastic Approximation Algorithms with unbounded Stochastic Information Delays - Theory and Applications

SPEAKER: Adrian Redder, Ph.D. student, Paderborn University, Germany ABSTRACT:At the advent of 2024, artificial intelligence (AI)-driven language and robotic systems are revolutionizing various domains and are expected to push the boundaries of human capabilities. The successes are based on advanced algorithms, but most importantly, on a growing consumption of computing resources. Training models with limited resources is therefore as important for reducing training time as it is for pushing the size of large AI models to new limits. This requires understanding the stability and convergence of learning algorithms, the most efficient use of parallel computing infrastructure, and how to adapt to errors caused by asynchronous, lightweight implementations.  To advance our understanding of these problems, I study distributed stochastic approximation algorithms. Such algorithms are defined by a coupled system of iterations that are adapted asynchronously by updates computed by a potentially large number of parallel computing resources. The framework jointly covers both asynchronous training of AI models as well as multi-agent learning in physically decentralized systems. The property that classifies these two scenarios as theoretically equivalent is that a set of variables is updated as a function of old versions of itself. In other words, the resulting iterations are affected by Age-of-Information (AoI). In this talk, I will first discuss general results on the stability and convergence of distributed stochastic approximation algorithms affected by AoI. I will then outline fundamental properties of AoI that enable these distributed stochastic approximation analyses, including an in-depth study of AoI arising from asynchronous parallel computing.  Everybody is welcome.