email@example.com Personal Site
Modern deep learning systems are so complex and nonlinear that many of their important features fall far outside our current mathematical understanding. Humans’ approach to understanding them has been almost like exploring a numerical universe whose rules we do not know, with researchers and practitioners often relying on trial and error and built-up intuition as opposed to principled math. My current research aims to build a more principled understanding of the numerical worlds of machine learning with tools physics uses to understand the real universe, including percolation theory, adiabatic evolution, and building tractable toy models. These approaches can yield testable predictions, so this work involves both theory and numerical experiments.
I’ve long been drawn to both physics and computer science. My first foray into research as an undergrad was designing low-level control schemes for quantum computers, and I spent a gap year in Sweden learning to nanofabricate superconducting quantum circuits. I loved the high dimensionality of quantum computing theory and the alien world of the very small one has to manipulate while nanofabricating, but doing a PhD in purely physics felt too narrow to me. I joined Berkeley as a physics PhD student in the fall of 2019 and grew interested in how neural nets distill patterns out of data, which led me to the Redwood Center.
Apart from science, I like making puzzles, running, playing music, and balancing things.