• 作者： 游寶達; 方光輝
• 作者服務機構： 中正大學資訊工程研究所
• 中文摘要： 堆疊類神經網路為一種基於正布林函數為非線性計算子之類神經網路，而正布林函數可以分類為四族一顯然、遞減、遞增及混合型四族。又分析堆疊類神經網路之行為中，以分析其“收斂行為”最為重要。目前已經證明出顯然、遞減及遞增型具有收斂性質，而混合型之收斂行為非常難分析。因此在本篇文章中，我們提出一族新的混合型堆疊類神經網路，稱之為圖形式堆疊類神經網路，其所基於之正布林函數可以利用一般的圖形來表示之。圖形式表示法的最大優點，在於其具有非常精簡之表示式，而且其收斂行為的分析也比較簡單。
• 英文摘要： The class of stack neural networks is a new class of neural networks whose nonlinear thresholdoperators are based on stack filters which can be classified into four different types-trivial, decreasing,increasing and mixed stack filters. One important concept used to characterize the retrieval operationof stack neural networks is that of the convergence property and it has been shown that all trivial, decreasingand increasing stack neural networks possess the convergence property; nevertheless, mixed stack neuralnetworks do not necessarily possess the convergence property. In this paper, we will investigate the convergence property of a large subclass of stack neural networksbased on the graph-based stack filters, which are a subclass of stack filters that can be easily expressedby undirected graphs. The graph-based representation provides a representation of the associative memoryof this new proposed neural networks. We will also investigate the convergence property of mixed graph-based stack filters with a new method proposed in this paper.
• 中文關鍵字： stack filter; stack neural network; praph-based representation; convergence property
• 英文關鍵字： --