摘要
本篇論文將探討非監督式學習 Self-Organizing Map (SOM) 演算法以下簡稱SOM，SOM跟其他 Artificial Neural Network (ANN)一樣是模仿人體神經網路的數學模型，但在設計上卻又跟其他ANN模型有所區別，SOM演算法使得與輸入向量相似的神經元彼此群聚以達到學習的成效，其中相似度是根據歐氏距離來計算，與輸入向量最靠近的神經元被稱作獲勝神經元或最佳匹配單元(BMU)，SOM中的BMU將根據輸入向量進行調整並靠近，直至SOM找不到任何可被更新的神經元，細節部份會在之後的章節進行討論，實驗結果將使用 8OX、Iris、Breast cancer Wisconsin (diagnostic) 以及HIV-1 protease cleavage data sets當作訓練集與其他傳統非監督式學習像是K-means、Agglomerative Hierar-chical Clustering及AdaBoost、Random Forest等監督式學習進行比較。
並列摘要
This thesis will discuss the unsupervised learning self-organizing map algo-rithm hereinafter referred to as SOM. Like other artificial neural networks (ANNs), SOM is a mathematical model that imitates the human neural network, but it is dif-ferent from other ANN models in terms of design. SOM algorithm makes neurons similar to the input vector cluster with each other to achieve the effect of learning. The similarity is calculated based on the Euclidean distance. The neuron closest to the in-put vector is called winner neuron or the best matching unit (BMU). The BMU will be adjusted and approached according to the input vector until SOM cannot find any neurons that can be updated. The details will be discussed in the following chapters. Experimental results on 8OX, iris, breast cancer Wisconsin (diagnostic), HIV-1 prote-ase cleavage data sets by SOM will be illustrated with other traditional clustering al-gorithms such as K-means, Agglomerative Hierarchical Clustering and supervised learning AdaBoost and Random Forest for a comparison.