Analysis and Training of Neural Networks Using Voronoi Diagrams and Graph Decomposition

本文提出了一种新的多层神经网络训练问题的解决方案-确定神经元的连接权-使用组合几何中一种非常强大的结构，称为Voronoi图。它还将解决同样重要的问题，即如何将神经网络划分为更小的子网络，这些子网络可以单独使用，也可以一起使用，以近似原始网络的功能。

This proposal presents a novel solution to the multilayer neural network training problem - determining the connection weights of the neurons - using a very powerful construct from combinatorial geometry called a Voronoi diagram. It will also address the equally important problem of how to partition a neural network into smaller subnetworks which can then be used either individually or together to approximate the function of the original network.