Physics ∩ ML

a virtual hub at the interface of theoretical physics and deep learning.

05 Oct 2022

Maths with transformers

François Charton, Meta

Transformers can be trained, from synthetic data, to solve problems of mathematics, by considering them as translations from problems into solutions. Models achieve high accuracy on a variety of tasks, learn deep mathematical properties, and generalize out of distribution if their training set is selected with care. I illustrate their use on several problems, from symbolic integration to numerical computation, and symbolic regression.