DDPS | “A first-principles approach to understanding deep learning” скачать в хорошем качестве

DDPS | “A first-principles approach to understanding deep learning”

DDPS Talk date: November 15th, 2024 Speaker: Yasaman Bahri (Google DeepMind, https://sites.google.com/view/yasaman...) Description: Recent years have seen unprecedented advancements in the development of ML and AI; for the sciences, these tools offer new paradigms for combining insights developed from theory, computation, and experiment towards design and discovery. Beyond treating them as black boxes, however, uncovering and distilling the fundamental principles behind how systems built using neural networks work is a grand challenge, and one that can be aided by ideas, tools, and methodology from physics. I will describe some of my past work that takes a first-principles approach to deep learning through the lens of statistical physics, exactly solvable models and mean-field theories, and nonlinear dynamics. I will first survey connections we discovered between large-width deep neural networks, Gaussian processes, and kernels, which bear application to data-driven research in the physical sciences. I’ll then discuss our work on understanding some facets of “scaling laws” describing the rate of improvement of an ML model with respect to increases in the amount of training data, model size, or computational resources. Bio: Yasaman Bahri is a Research Scientist at Google DeepMind with research interests at the interface of statistical physics, machine learning, and condensed matter physics. In recent years, she has worked on the foundations of deep learning from a physics perspective. She has been an invited lecturer at the Les Houches School of Physics and was a co-organizer of a recent program at the Kavli Institute for Theoretical Physics. Previously, she received a Ph.D. in Physics at UC Berkeley. DDPS webinar: https://www.librom.net/ddps.html 💻 LLNL News: https://www.llnl.gov/news 📲 Instagram:   / livermore_lab   🤳 Facebook:   / livermore.lab   🐤 Twitter:   / livermore_lab   About LLNL: Lawrence Livermore National Laboratory has a mission of strengthening the United States’ security through development and application of world-class science and technology to: 1) enhance the nation’s defense, 2) reduce the global threat from terrorism and weapons of mass destruction, and 3) respond with vision, quality, integrity and technical excellence to scientific issues of national importance. Learn more about LLNL: https://www.llnl.gov/. IM release number is: LLNL-VIDEO-871423