MICS Seminar: Aurélien Sagnier

In this talk, we will try to illustrate what topossic ideas on learning could give on very simple structures like graphs and how one could actually work and compute handling those structures. It was achieved thanks to the formalism of cellular sheaves and sheaf-diffusion as developed by R.Ghrist and his school (Curry, Hansen, Riess,…) and has already been used in practice by this school and the group of M. Bronstein (Bodnar, Di Giovanni, Chamberlain, Barbero, Lio,…).
In this talk, I will first recall this formalism, then show an application of it to a very efficient solution of the problem of grammar learning studied in Lake and Baroni “Human-like systematic generalization through a meta-learning neural network” Nature 623, 115-121 (2023) and finally open on some topossic prospects for machine learning.