¾ÅÉ«ÊÓƵ

Event details

Title: Next-Generation GW Inference: Traditional Bayesian Methods in an AI Era

Abstract: Over a decade on from the monumental first detection, gravitational waves are rapidly transforming how we probe black hole formation, cosmic expansion, and fundamental physics. This rich science has relied on extracting the source properties of signals, a task which requires substantial computational effort and days to weeks of time. With the LVK O5 observing run on the horizon, and the advent of future space- and ground-based detectors, the next decade of signals will be louder, longer, and more numerous, magnifying the challenges of data analysis. In response to this, the community has begun to leverage machine learning and AI at scale. Concurrently, the explosion of AI has driven a shift in the computing landscape, with Graphics Processing Units (GPUs) playing an increasing role in scientific research. Indeed, the most promising machine learning methods in our field make use of this architecture. Although it has become natural to associate GPUs almost exclusively with neural networks, in this talk I will ask a different question: what happens if we put more conventional, interpretable statistical methods onto this same hardware? After providing an overview of gravitational-wave parameter estimation, I will demonstrate how community-standard Bayesian sampling algorithms can be heavily accelerated using GPUs. I will argue that, to successfully achieve next-generation inference, the algorithmic advancements of machine learning and the hardware advancements of GPUs should be viewed as distinct yet complementary tools. This approach redefines what it means to perform science in an AI era, demonstrating that traditional statistical methods still have a scalable, vital role to play alongside emerging machine learning techniques, ultimately allowing us to develop unique hybrid approaches to answer our most challenging questions.