Learning 3-manifold triangulations (Q7004128)

In this article, the authors evalute the performance of modern machine learning tools on the problem of recognizing 3-manifolds and knot complements. The starting point of their analysis is to choose a small collection of manifolds of interest and generate a large number of triangulations of these manifolds using Pachner moves. They first provide an analysis of the obtained Pachner graphs from the perspective of network science: centrality, Wiener index, clustering coefficients, etc., leading them to an intriguing conjecture connecting the growth rate of the Pachner graph of a cusped hyperblic 3-manifold to the inverse of the systole.\N\NThen, these triangulations are used to train and evaluate the performance of neural networks at recognizing 3-manifolds from their triangulations, and more precisely from the isomorphism signatures of their triangulations, which are very concise string encodings of their gluing structures. Extensive results are provided, highlighting in particular that some pairs of manifolds are harder to distinguish than others, and that the transformer architecture, despite being notably well-tailored to string problems, does not perform well on the task at hand.