本文選題：卷積結(jié)構(gòu) + 獨(dú)立主成分分析　； 參考：《南昌大學(xué)》2017年碩士論文

【摘要】：深度學(xué)習(xí)是目前研究的熱點(diǎn),本文基于前人的研究,針對(duì)傳統(tǒng)卷積神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)算法在卷積核數(shù)量確定、卷積核參數(shù)初始化以及網(wǎng)絡(luò)層與層之間通道選擇不能夠自主學(xué)習(xí)的問題,提出了基于獨(dú)立主成分分析和基于云模型的卷積神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)訓(xùn)練方法。首先分析研究主成分分析的主向量與卷積神經(jīng)網(wǎng)絡(luò)卷積核數(shù)量之間的關(guān)系;然后運(yùn)用獨(dú)立主成分神經(jīng)網(wǎng)絡(luò)對(duì)無標(biāo)記圖像混合信號(hào)分析,確定深度卷積核數(shù)量,模仿快速固定算法對(duì)卷積核參數(shù)初始化;最后采用云模型選擇器對(duì)模型層與層之間通道連接進(jìn)行選擇,從而達(dá)到減少網(wǎng)絡(luò)訓(xùn)練步數(shù)的目的。實(shí)驗(yàn)結(jié)果表明通過主成分分析所得抽樣樣本的平均主向量個(gè)數(shù)近似模型在測試集準(zhǔn)確率基本不變的情況下卷積核數(shù)量的拐點(diǎn),減少了重復(fù)實(shí)驗(yàn)來確定卷積核數(shù)量的次數(shù)。獨(dú)立主成分析以及云模型選擇器使得模型訓(xùn)練步數(shù)減少,加快了模型收斂,所提算法要比經(jīng)典算法運(yùn)行速率上提高5.4%左右,測試集準(zhǔn)確率基本與原卷積神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)基本持平。
[Abstract]:Deep learning is a hot topic at present. Based on the previous research, this paper proposes the problem that the number of convolution kernel is determined, the parameter of convolution kernel is initialized and the selection of the channel between layer and layer can not be autonomous learning. A convolution neural network based on independent principal component analysis and cloud model is proposed. The relationship between the principal vector of the principal component analysis and the convolution kernel number of the convolution neural network is analyzed. Then the independent principal component neural network is used to analyze the unmarked image mixed signal, to determine the number of the deep convolution kernel, and to initialize the convolution kernel parameters by the fast fixed algorithm. Finally, the cloud model is used to select the model. The selector chooses the channel connection between the model layer and the layer to reduce the number of network training steps. The experimental results show that the average principal vector number approximation model of the sample sample obtained by the principal component analysis reduces the volume of the volume of the kernel number under the condition that the accuracy of the test set is basically the same, and the repeated experiment is reduced to determine the volume. The number of kernel number, independent principal analysis and cloud model selector reduce the number of model training steps and accelerate the convergence of the model. The proposed algorithm is about 5.4% higher than the classical algorithm, and the accuracy of the test set is basically equal to the original convolution neural network structure.
【學(xué)位授予單位】：南昌大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2017
【分類號(hào)】：TP18

【參考文獻(xiàn)】

相關(guān)期刊論文 前10條

1 李建林;;一種基于PCA的組合特征提取文本分類方法[J];計(jì)算機(jī)應(yīng)用研究;2013年08期

2 孫志軍;薛磊;許陽明;王正;;深度學(xué)習(xí)研究綜述[J];計(jì)算機(jī)應(yīng)用研究;2012年08期

3 李宏明;;基于多元統(tǒng)計(jì)分析的地方高校課堂教學(xué)質(zhì)量評(píng)價(jià)——以臺(tái)州學(xué)院為例[J];臺(tái)州學(xué)院學(xué)報(bào);2010年03期

4 ;Study on the Universality of the Normal Cloud Model[J];Engineering Sciences;2005年02期

5 李德毅,劉常昱,杜瀊,韓旭;不確定性人工智能[J];軟件學(xué)報(bào);2004年11期

6 劉常昱,馮芒,戴曉軍,李德毅;基于云X信息的逆向云新算法[J];系統(tǒng)仿真學(xué)報(bào);2004年11期

7 夏長亮,文德,范娟,楊曉軍;基于RBF神經(jīng)網(wǎng)絡(luò)的無刷直流電機(jī)無位置傳感器控制[J];電工技術(shù)學(xué)報(bào);2002年03期

8 蔣嶸,李德毅,范建華;數(shù)值型數(shù)據(jù)的泛概念樹的自動(dòng)生成方法[J];計(jì)算機(jī)學(xué)報(bào);2000年05期

9 邸凱昌,李德毅,李德仁;云理論及其在空間數(shù)據(jù)發(fā)掘和知識(shí)發(fā)現(xiàn)中的應(yīng)用[J];中國圖象圖形學(xué)報(bào);1999年11期

10 金忠,胡鐘山,楊靜宇;基于BP神經(jīng)網(wǎng)絡(luò)的人臉識(shí)別方法[J];計(jì)算機(jī)研究與發(fā)展;1999年03期

相關(guān)碩士學(xué)位論文 前1條

1 王華明;基于多向獨(dú)立主成分分析方法的間歇過程監(jiān)控研究[D];中國石油大學(xué);2010年

，

本文編號(hào)：2030980