 |
Yasaman Bahri
|
I am a scientist with broad interests in machine learning (theory and empirics), particularly in the foundations of the field. My work tends to sit at a multidisciplinary interface of learning, physics, statistics, and computation. A common theme is to better understand mechanisms behind how things work and, where possible, leverage this towards challenging design problems. I am also interested in the use of machine learning in the physical sciences, in particular condensed matter & materials science.
A sample of topics from some of my recent work in machine learning include:
Theoretical and empirical understanding of the role of scale in deep learning (‘‘scaling laws")
Exact connections between neural networks, Gaussian processes, and kernel methods
Phase transitions and the dynamics of gradient descent in deep learning
Distribution shift
I was trained as a theoretical quantum condensed matter physicist, and I received my Ph.D. in Physics from UC Berkeley in 2017. My graduate work is specifically in the field of quantum many-body theory and strongly correlated systems. I was fortunate to have Professor Ashvin Vishwanath as my thesis advisor. My scientific interests have always been broad, and I worked on several different areas as part of my doctoral research, including topological phases, many-body localization, and non-Fermi liquids. My dissertation proposed new classes of quantum behavior; new routes towards realizing exotic quantum phases; and new classes of mechanical behavior through topological mechanisms. I got started in theory as an undergraduate through research on tensor networks and entanglement in quantum systems with Professor Joel Moore at UC Berkeley, which was also the subject of my honors senior thesis.
Recent News
Gave a guest lecture in CS 159 at Caltech on theoretical aspects of deep learning.
I am co-organizer for the ICML 2021 workshop, Overparameterization: Pitfalls & Opportunities.
We have posted a new preprint: ‘‘Explaining Neural Scaling Laws."