臺灣博碩士論文加值系統

本論文的目的在於針對一般多層類神經網路學習的問題, 提出解決的方法
。一般多層神經網路的學習演算法為回傳式學習法。這個演算法有速度太
慢, 網路大小難決定, 及落至局部最小的缺點。我們分別提出了三個不同
的方法, 來解決這些問題。這三個方法分別為: 1.給定初值 2.各個神經
元分別負責一度空間的對應 3.利用保存鄰近關係的方法。這些方法可以
應用在圖形辨認上。圖形辨認的問題中, 圖形資料的分布主要有兩類, 一
為具有天然的相似性, 一為不具自然的相似性。我們的方法對具有天然的
直線狀分布或曲線分布的資料分類有大幅改善。

The objective of this research is to propose methods on solving
the learning problems of multi-layer neural networks. The most
well-known and commonly used learning algorithm is Back
Propagation (BP) algorithm. There are three main drawbacks of
BP: 1. the slowness of the learning speed, 2. the convergence
to local minima, and 3. the absence of any theoretical result,
allowing for a priori determination of an optimal network
architecture for a given task. To solve these problems, we
propose three methods: The first is to initialize weights in
multi-layer quadratic sigmoid networks; The second is to learn
in successive residual space; The third is using the topology
preserving maps formmed in MLPs. These methods can be applied
to pattern recognition problems espically when the training
patterns have lined structure.