Скачать или смотреть Computational Aspects of Equivariant Neural Networks - Risi Kondor

Equivariant networks have become the de-facto standard architecture for modeling physical systems with neural networks from learning force fields for molecular dynamics to learning the wave function of quantum systems. In addition to operations encountered in neural nets, equivariant networks involve a couple of additional operators, specifically, the Clebsch-Gordan product of the underlying symmetry group. In practical implementations these operators often become the computational bottleneck and ultimately constrain the expressive power of the entire network. In this talk, I will describe a combination of mathematical and computational approaches for accelerating equivariant architectures.

Pour vous inscrire au prochain : https://ivado.ca/evenements/seminaire...