目前,機(jī)器學(xué)習(xí)正在蓬勃發(fā)展。機(jī)器學(xué)習(xí)不僅與更快,更容易,更便宜的數(shù)據(jù)的收集與處理的方法有關(guān),還與來(lái)自于物理學(xué)、生物學(xué)、經(jīng)濟(jì)等學(xué)科采集的數(shù)據(jù)進(jìn)行建模的方法的發(fā)展有關(guān)。在一些任務(wù)中,當(dāng)難以建模時(shí),可運(yùn)用深度學(xué)習(xí)方法,使用各種線性,非線性轉(zhuǎn)換(通常表示為人工神經(jīng)網(wǎng)絡(luò))對(duì)數(shù)據(jù)進(jìn)行抽象建模,神經(jīng)網(wǎng)絡(luò)已成功用于解決諸如預(yù)測(cè)、模式識(shí)別、數(shù)據(jù)壓縮等問(wèn)題。本文研究的目的旨在對(duì)一些深度學(xué)習(xí)庫(kù)的軟件工具進(jìn)行比較分析,如:Caffe、Pylearn2、Torch和Theano。本研究是以解決深度學(xué)習(xí)的疑難任務(wù)之一即識(shí)別手寫數(shù)字的問(wèn)題為例進(jìn)行的,以手寫數(shù)字圖像數(shù)據(jù)庫(kù)(MNIST)用作測(cè)試數(shù)據(jù)集。本文中,將對(duì)兩種類型的神經(jīng)網(wǎng)絡(luò)MLP和CNN進(jìn)行CPU和GPU研究分析,研究的結(jié)果通過(guò)六個(gè)方面對(duì)上述四個(gè)深度學(xué)習(xí)庫(kù)進(jìn)行評(píng)估:1)學(xué)習(xí)速度;2)分類速度;3)易用性;4)配置靈活性;5)功能性;6)文檔的可訪問(wèn)性、易用性。根據(jù)上面研究的結(jié)果,本文還討論了使用其中某些庫(kù)來(lái)解決尋找行人和汽車的問(wèn)題。

【文章頁(yè)數(shù)】：66 頁(yè)

【學(xué)位級(jí)別】：碩士

【文章目錄】：
摘要
Abstract
Introduction
1 Methods of deep learning
    1.1 Unsupervised learning systems
        1.1.1 Limited Boltzmann machine
        1.1.2 Autocoder
        1.1.3 Deep belief network
    1.2 Teaching with the teacher
        1.2.1 The convolutional neural network
        1.2.2 Artificial Neural Networks
2 Research purpose
    2.1 MNIST Dataset
    2.2 Software tools for solving problems of deep learning
        2.2.1 Operation system
        2.2.2 CUDA Toolkit
        2.2.3 Caffe library
        2.2.4 Pylearn2 library
        2.2.5 Torch Library
        2.2.6 Theano library
3 Setting up libraries
    3.1 Setup Caffe library
    3.2 Setup Pylearn2 library
    3.3 Setup Torch library
    3.4 Test infrastructure
    3.5 Network topologies and learning parameters
4 Experimental results
    4.1 Library Evaluation Criteria
    4.2 Scoring and comparing libraries
5 Conclusions and Future Work
    5.1 Conclusions
    5.2 Future Work
Acknowledgements
References
Academic Achievements



本文編號(hào)：3739382