- 作者: 曾怜玉; 黃子漢
- 作者服務機構: 中興大學應用數學研究所
- 中文摘要: 所有前饋式類神經網路皆適合在監督學習後當做模式分類器。在前饋式類神經網路中,後向傳播學習法則是一種功能很強,且已經成功應用到各種不同的問題上。然而,因後向傳播學習法則是一種“黑箱”學習法,故仍然存在某些缺點,特別是學習時問往往比較長、當學習之樣本數量太大時,學習成功率不是很理想,以及隱藏單元數目無法預知等缺點。因此,Pao提出函數鏈路類神經網路(Functional-Link Neur al Network)來減少隱藏層和增快學習速度。但仍存在「如何求取函數」之缺點。因此,我們提出利用網路分割特性和交換邏輯理論之方法,能設計一「邏輯函數類神經網路」以增進前饋式類神經網路之學習速度。由我們的實驗結果可知,前饋式類神經網路不但可擺脫傳統之盲目學習,而且我們提出之邏輯函數類神經網路還可獲得下列優點:網路變小、學習成功率提高、以及學習速度變快。
- 英文摘要: Feedforward networks are good pattern classifiers and can be used with supervised learning. Amongthe feedforward network learning rules, the backpropagation learning rule is a powerful one which hasbeen successfully applied to a variety of problems. However, the process of backpropagation learningis, somewhat, "blind learning". It suffers from several limitations; in particular, the learning speed isoften unacceptably low, the optimum number of hidden nodes is not known in advance, so learning maynot succeed if there are a large number of input patterns. Therefore, a functional-link neural networkwas proposed by Pao to eliminate hidden layers and improve the speed of learning. However, the drawbackof such a network is the combinatorial explosion of high-order terms. In this paper, we propose a LogicFunction Neural Network (LFNN). An input-output relation with multiple output nodes is first dividedinto several input-output relations, each with exactly one output node, and then a logic function neuralnetwork is designed by applying logic design theory to help network learning. The advantages of theproposed neural network are: the network size is smaller, the successful rate of learning is higher, andlearning is faster as well.
- 中文關鍵字: feedforward network; backpropogation learning rule; logic function; pantitioned netword
- 英文關鍵字: --