Marco Pacini
Researcher at Fondazione Bruno Kessler, Trento, Italy
My research focuses on the mathematical foundations of Geometric Deep Learning and Equivariant Machine Learning, with two main directions:
-
Approximation Theory of Equivariant Neural Networks. A substantial part of my work concerns density and approximation properties of equivariant neural networks, with particular attention to the role of their separation power.
-
Characterization of Equivariant Activations. Since point-wise nonlinearities such as ReLU are not compatible with arbitrary symmetries, part of my work characterizes which pairs of group representations and scalar activation
functions induce equivariant nonlinearities.
