Deep Learning from the Perspective of a Physicist

In one of many worlds, we will either (i) discuss a way to minimally extend the traditional workhorse of machine learning, the linear model, by considering the quadratic model, an interacting model that exhibits feature learning, (ii) present a solvable model of neural scaling laws that we can understand using tools from random matrix theory, or (iii) give pedagogical introduction to "language models" and the transformer architecture that powers exciting AI systems such as ChatGPT. (We probably live in the world described by the third item.)