- 作者: 游寶達; 張政德
- 作者服務機構: 中正大學資訊工程研究所
- 中文摘要: 堆疊類神經網路是一種基於堆疊濾波器為其非線性計算子之類神經網路。整體而言,欲設計較佳的堆疊類神經網路,必須使錯誤記憶(false memory)愈少愈好。此外,堆疊類神經網路對訊號之間的離散計算行為,必須詳加分析。因此,在本論文中,兩訊號在堆疊類神經網路中的離散計算行為被詳細分析之,我們提出S-cut及RS-cut等方法,以作為分析的工具。另外,在堆疊類神經網路中,我們也分析兩訊號間的多步計算模式,其中因為計算複雜度的關係,我們僅詳細分析4-cube之行為。
- 英文摘要: Stack neural networks are a new class of neural networks which are based on stack filters (Wendt,1986) as their nonlinear operation. In our previous research, we have discussed some preservation behaviorsof stack neural networks (Yu, 1993). However, the global behaviors of preservation behaviors (learningbehaviors) and discrete computation (retrieval) of stack neural networks are not completely understoodup to now. In this paper, we design several algorithms to investigate the behaviors of the discrete iterationof stack neural networks. We consider a pair of signals as a training pair such that the desired stackneural network can be trained. If the stack neural network with this training pair can't be trained in onetraining step, then we want to find the multi-step training behaviors of this stack neural network. Hence,we propose two algorithms to find all possible input and output signals for which they can form one-step training pairs of stack neural networks. These methods don't need to directly use any stack filterto compute the output and can then get the same results as filtering by all positive stack filters. On theother hand, we can use these algorithms to determine the behaviors of the discrete iteration of stack neuralnetworks for all signals. According to several induction and experimental results, we also find somebehaviors of discrete computation for general k steps in this paper.
- 中文關鍵字: stack filiters; positive Boolean function; S-cut; RS-cut; stacking property; positive extension; negative extension
- 英文關鍵字: --