Kähler Geometry of Quiver Varieties and Machine Learning

Abstract

We develop an algebro-geometric formulation for neural networks in machine learning using the moduli space of framed quiver representations. We find natural Hermitian metrics on the universal bundles over the moduli which are compatible with the GIT quotient construction by the general linear group, and show that their Ricci curvatures give a Kähler metric on the moduli. Moreover, we use toric moment maps to construct activation functions, and prove the universal approximation theorem for the multi-variable activation function constructed from the complex projective space.

v2: expanded on related works. 47 pages, 13 figures

References in corpus (5)

• Ricci flow with surgery on three-manifolds
• Identifying and attacking the saddle point problem in high-dimensional non-convex optimization
• On the saddle point problem for non-convex optimization
• Covariance in Physics and Convolutional Neural Networks
• Graph Laplacians, Riemannian Manifolds and their Machine-Learning