(Maths-ia) From fitting ellipsoids to random points, to learning in large neural networks

I will first discuss a problem of high-dimensional probability known as the ellipsoid fitting conjecture. I will present recent progress on this conjecture using both non-rigorous analytical tools of statistical physics, and rigorous methods based on the combination of universality results in statistics and extensions of classical approaches in random convex geometry. In a second part, I will discuss how the techniques developed to analyze ellipsoid fitting can be used to sharply characterize optimal learning in a wide neural network with a quadratic activation function, as well as in a model of learning from long sequences of high-dimensional tokens.