線上自我調節特徵映射神經網路與應用

Abstract

在本篇論文中，我們提出了一個新的類神經網路－自我調節特徵映射網路。此網路我們發展了一套含有自我調節機制的非監督式學習演算法。經過訓練過後的代表性神經元相互連結所組成的網路結構不僅顯示的輸入資料分佈的空間關係，也能很好的量化資料。自我調節特徵映射網路可自動將有連接的神經元分離成各個集合，而集合的數目也就代表了學習過程中所該使用的群聚數目。自我調節機制的概念在於結合數學上的統計、神經學上的用進廢退與基因學上的突變。每一個代表性的神經元在訓練的過程中都分成三個階段：遷移、分化和衰退。代表性神經元在遷移階段會各自移動到適合的位置。在分化階段的前期代表性神經元會根據輸入資料的局部密度來決定是否需要做分裂。然後在分化階段的後期會依照輸入資料的統計不斷調整代表性神經元與隣近神經元間在拓撲上的鍵結狀態。最後在衰退階段會不必要的代表性神經元將會被融合或刪除。除此之外，為了能處理非穩態資料進而達到線上學習，我們亦在神經元的學習率添加了突變機制。在實驗中，我們利用自我調節特徵映射網路處理一些在傳統非監督式學習網路(如自我組織特徵映射網路)無法處理良好的案例。更進一步，我們也採用幾項現實上的例子來表現出自我調節特徵映射網路的顯著特點。

In this thesis, we propose a novel artificial neural network, called self-adjusting feature map (SAM) network, and develop its unsupervised learning algorithm with self-adjusting mechanism. The trained network structure of representative connected neurons not only displays the spatial relation of the input data distribution but also quantizes the data well. SAM can automatically isolate each set of connected neurons, in which the number of the set may indicate the number of clusters to be used. The idea of self-adjusting mechanism is based on combining of mathematical statistics, neurological advance and retreat of waste, and mutation in genetics. For each representative neuron, there are three periods, migration, differentiation and decline, in its training process. The representative neurons in the migration period will make their own move to the appropriate position. The representative neuron will decide whether it needs to do a division according to the local density of the input data in the early stage of the differentiation period. Then it will adjust neighborhood neuron pair’s connected or disconnected topology constantly according to the statistics of input data in the late of differentiation period. Lastly the unnecessary representative neurons will be merged or deleted in the decline period. Besides, we also add the mutation in the learning rate of the neurons to deal with nonstationary data, and even online learning problem. In the experiments, we exploit SAM to handle some peculiar cases that cannot be well dealt with by classical unsupervised learning networks such as self-organizing feature map (SOM) network. Furthermore, we also take several real world cases to exhibit the remarkable characteristics of the SAM