多層ニューラルネットワークに対する解の必要条件の定式化

概要

論文の詳細を見る
This paper formulates a necessary condition for unit's weights and bias to become the elements of solutions for multilayer nets in terms of normal vectors orthogonal to separation hyperplanes. Then it depicts the distribution of normal vectors of two-dimension satisfying the condition, and elucidates that the distributing domain for unipolar units activating from 0 to 1 is significantly different from that for bipolar units activating from -0.5 to 0.5. At the same time, it depicts the initial distribution of normal vectors with the weights and biases given ordinarily by random numbers with zero mean. A comparison of this distribution to that satisfying the necessary condition leads to the fact that the nets composed of the bipolar units are superior in convergence of the back propagation learning to those composed of the unipolar units.
近畿大学の論文
1996-12-01