斯坦福深度学习课件3 深度卷积神经网络的谐波分析bolcskei-stats385-slides

《斯坦福深度学习课件3 深度卷积神经网络的谐波分析bolcskei-stats385-slides》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、HarmonicAnalysisofDeepConvolutionalNeuralNetworksHelmutB}olcskeiDepartmentofInformationTechnologyandElectricalEngineeringOctober2017jointworkwithThomasWiatowskiandPhilippGrohsImageNetImageNetskirockplantcoeeImageNetskirockplantcoeeCNNswintheImageNet2015challenge[Heetal.,2015]Desc

2、ribingthecontentofanimageCNNsgeneratesentencesdescribingthecontentofanimage[Vinyalsetal.,2015]DescribingthecontentofanimageCNNsgeneratesentencesdescribingthecontentofanimage[Vinyalsetal.,2015]Carlos."DescribingthecontentofanimageCNNsgeneratesentencesdescribingthecontentofanimage[V

3、inyalsetal.,2015]CarlosKleiber."DescribingthecontentofanimageCNNsgeneratesentencesdescribingthecontentofanimage[Vinyalsetal.,2015]CarlosKleiberconductingthe."DescribingthecontentofanimageCNNsgeneratesentencesdescribingthecontentofanimage[Vinyalsetal.,2015]CarlosKleiberconducting

4、theViennaPhilharmonic's."DescribingthecontentofanimageCNNsgeneratesentencesdescribingthecontentofanimage[Vinyalsetal.,2015]CarlosKleiberconductingtheViennaPhilharmonic'sNewYear'sConcert."DescribingthecontentofanimageCNNsgeneratesentencesdescribingthecontentofanimage[Vinyalsetal.,2

5、015]CarlosKleiberconductingtheViennaPhilharmonic'sNewYear'sConcert1989."Featureextractionandclassicationinput:f=non-linearfeatureextractionfeaturevector(f)linearclassier(hw;(f)i>0;)Shannonoutput:hw;(f)i<0;)vonNeumannWhynon-linearfeatureextractors?Task:Separatetwocategoriesofd

6、atathroughalinearclassier1:hw;fi>0:hw;fi<0Whynon-linearfeatureextractors?Task:Separatetwocategoriesofdatathroughalinearclassier1:hw;fi>0:hw;fi<0notpossible!Whynon-linearfeatureextractors?Task:Separatetwocategoriesofdatathroughalinearclassierkfk(f)=11:hw;fi>0:hw;(f)i>0:hw;fi<

7、0:hw;(f)i<01notpossible!possiblewithw=1Whynon-linearfeatureextractors?Task:Separatetwocategoriesofdatathroughalinearclassierkfk(f)=1)isinvarianttoangularcomponentofthedataWhynon-linearfeatureextractors?Task:Separatetwocategoriesofdatathroughalinearclassierkfk(f)=1)isi

8、nvarianttoangularcomponent