頁籤選單縮合
題名 | 建構區塊型式的神經網路 |
---|---|
作 者 | 陳木松; | 書刊名 | 大葉學報 |
卷期 | 2:1 1993.12[民82.12] |
頁次 | 頁97-108 |
分類號 | 312.2 |
關鍵詞 | 建構區塊; 多層認知器的神經網路; 泰勒多項式; 映射; 激發函數; Building blocks; Multi-layer perceptron neural networks; Taylor series; |
語文 | 中文(Chinese) |
中文摘要 | 本論文是針對如何以模組化的連構區塊(Modular Building Blocks)來設計多層認知器的神經綱路(Multi-layer Perceptron Neural Networks)。基本上,所謂模組化建構區塊的神經網路的構想理念,是將輸入至神經網路的多項式函數分解成 k 次方項(k-degree Monomial)。每一個別項再分別由模組化的建構區塊實現得到所要的輸出。再將所有個別子網路或建構區塊輸出的結果整合,而得到崳入多項式函數的映射(Mapping)結果,亦即y=f(x)。如果輸入多項式函數為已知,則模組化建構區塊的設計方法將可決定所需神經網路的架構,及神經元相互間的連接權值。本文將討論如何以泰勒多項式(Taylor Polynomi.al)型式的激發函數(Activation Functions)設計k次方項的建構區塊,並推論相關的數學模式及如何控制其準確度等。最後,再以實例設計加以詳述之。 |
英文摘要 | The major objective of this paper is to design the multi-layer perceptron neural networks using modular building blocks. The basic concept of modular building blocks is to decompose the input polynomial function of the network into k-degree monomial. Then, every term of the polynomial function is mapped to the modular subnets. Finally, all the modular subnets are assembled together to form the desired functions. It is assumed that the activation function is approximated by the truncated Taylor series. Related mathematical model and theorems are introduced to control the accuracy of the approximation. The principle advantages to this building blocks approach are (1) the network topology is known if the conventional algorithm is completely described and (2) the mapped network has good initial weights. Examples are given to illustrate the design procedures and demonstrate the validity of the proposed mapping algorithms. |
本系統之摘要資訊系依該期刊論文摘要之資訊為主。