基於多層自我組織映射圖之可視覺化深度學習模型
摘要
隨著科技的發展,對於大腦的研究也越來越豐富,認為大腦的結構是具有學習能力,並想試著學習模仿大腦的結構。如果我們可以將機器加入學習能力,學習生活周遭的事情,是否能夠利用機器取代人類,做一些簡單、重複性高的事,讓人類的生活更加輕鬆。 本論文所使用的網路架構是結合非監督式和監督式的類神經網路。透過影像的輸入,將影像結合並傳到下一層皮質區,最後傳達大腦時,由大腦的經驗做判斷。本論文架構採用自我組織映射,將特徵相似度高的放在一起;再透過適應共振理論,對於整個架構的輸出結果做儲存;最後,利用學習向量量化網路,對適應共振理論所儲存的特徵映射響應圖進行微調,而非使用一般常見的梯度修正。另外,我們透過可視覺化,將每層的特徵轉換成具有解讀性特徵影像,使人可以有效的解讀特徵所賦予的意義。 本論文的實驗中,共使用兩種資料集,分別為手寫辨識資料集和歌曲資料集,並對此架構中的特徵映射圖、適應共振理論的警戒參數、影像二值化、影像結合進行比較和結果分析,在最後對特徵圖做可視覺化的呈現。 關鍵字:視覺皮質、自我組織映射、適應共振理論

關鍵字: Kinect、視覺障礙者、卷積神經網路、物品偵測

 

 

A Visualized Deep Learning Model Based on Multilayer Self-Organization Maps
Abstract
In recent years, many people try to understand the structure of the brain. We believed that if we can add machines to learning. It is possible to use machines instead of humans and do some simple, repetitive things. This neural network architecture is a combination of supervised and unsupervised neural networks. This paper uses the k-means algorithm to group the input image. And then use the self-organizing map to generate feature map. And then use adaptive resonance theory to save the results of response map. Finally, use the learning vector quantization network fine-tuning the results of adaptive resonance theory without the use of Gradient. In addition, we can visualize and transform each layer of features into images that can be understood by the human’s eyes. In this paper’s experiments, we use two types of datasets. One is MNIST, another is song datasets. The experiment includes the feature maps, the alert parameters of the adaptive resonance theory, binary of image, and image combination. In the end, the feature is visualized presented The experiment in this thesis has compared to the feature maps, the alert parameters of ART, the image binary, and the different method of the image’s merger. Keywords: Visual Cortex, Self-organizing Map, Adaptive Resonance Theory