Machine Learning in Physics

I taught this class at the TU Wien in Spring 2021. It is an introduction to basic machine learning techniques, with extensive formal justification, and applied to simple problems from physics.

After completing the class, students are supposed to be able to take a physics problem and:



The lecture can be roughly divided into three parts:
  1. Python and Optimziation (exercises 1–3): Since physics undergrads at the TU Wien learn coding in C++ and not Python, the class starts with a Python crash course. As exercises that go together with this intro and also are relevant to ML, we are looking at simple optimization problems and solve them with (accelerated) gradient descent or Newton's method.
  2. Linear models (exercises 4–7): This is the centrepiece of the class. I am introducing the “scientific method for machines”, i.e., the machine learning workflow, and we are going to run through it multiple times while cranking up the complexity. The core analysis technique used is the singular value decomposition. We use it to prove the bias–variance tradeoff and perform convergence analysis for gradient descent.
  3. Advanced models (exercises 8–11): Here we construct artificial neural networks by building them up from logistic models, which are introduced first. We then move to unsupervised learning, to make sure that students have seen another paradigm than supervised learning. We can reuse our SVD analysis to construct low-rank approximations, which concludes the class.


The idea of this class is a little different from your usual shiny happy pictures class on machine learning: there is a heavy focus on simple models, in particular the linear model, and rigorous analysis.

There are three reasons for this:

  1. The linear model can be used to justify a lot of the “common wisdom” in machine learning rigorously and at same time pedagogically. I found that actually being able to understand and prove, e.g., convergence behaviours is as important as observing it in-the-wild.
  2. From my experience, students approach a machine learning problem in physics from “the wrong end”, i.e., by throwing the most sophisticated model at it without much a-priori analysis instead of doing sophisticated a-priori analysis and then start with the simplest model that is reasonable.
  3. The convex cost functions means linear and logistic models are straight-forward to train. I found writing simple training codes yourself greatly helps “demystifying” machine learning libraries.