Date and Time
Location
Abstract
The remarkable success of modern AI has far outpaced our understanding of the fundamental principles and mechanisms that govern and drive it. At the same time, a deeper understanding of representation learning - how representations are best learned, shaped, and controlled - is important as data-driven methods become part of the fabric of science. In this talk, I’ll highlight some of my past work in the ‘physics of AI,’ showing that a deeper understanding of the principles behind AI is possible when treating it as a complex physical system. I’ll give examples in which I’ve developed new theories, insights, and methods for learning with neural networks and interpretability of large AI models as a result. As one recent illustration, I’ll explain how a universal geometry in representational manifolds within large language models emerges from symmetry. Looking forward, I’ll also discuss opportunities in the complementary direction, on how machine learning and AI can advance science, with a particular focus on the realm of materials and quantum phenomena.
Bio
Yasaman Bahri is a Research Scientist at Google DeepMind with research interests at the intersection of physical science and AI. Her past research has laid foundations for the emerging field of the physics of learning and neural computation, and she serves as an affiliate of a Simons Foundation Collaboration on the same theme. Originally trained as a theoretical quantum condensed matter physicist, she is also interested in developing methods for and analyzing representation learning for materials, molecules, and quantum systems. She received her Ph.D. in Physics from UC Berkeley as an NSF Graduate Fellow.
Links
Poster
