Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing...
Transcript of Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing...
![Page 1: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/1.jpg)
This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under contract DE-AC52-07NA27344. Lawrence Livermore National Security, LLC
GraphSignalProcessingusingDeepNeuralNetworks
JayaramanJ.Thiagarajan
![Page 2: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/2.jpg)
2
JointWorkWith
Rushil AnirudhLLNL
BhavyaKailkhuraLLNL
PrasannaSattigeriIBMResearch
Karthi RamamurthyIBMResearch
![Page 3: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/3.jpg)
3
GraphsareEverywhere
Graphics Social Networks Medicine
Graphscancapturecomplexrelationalcharacteristicsforanalysisandinference
![Page 4: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/4.jpg)
4
UndirectedGraphs– BasicNotations
![Page 5: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/5.jpg)
5
SignalsonaGraphcanbeasInterestingastheGraphItself
Analyzecharacteristicsofthesignalbyexploitingthegraph
structure
§ Examples:— BrainnetworksandfMRIsignals— Generegulatorynetworksandgeneexpressionlevels— Socialnetworksandinformationcascades
![Page 6: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/6.jpg)
6
ConnectiontoFourierAnalysis– EigenvectorsasFrequencies
![Page 7: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/7.jpg)
7
§ UsingtheeigenvectorsoftheLaplacian,wecandefine
GraphFourierTransform
![Page 8: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/8.jpg)
8
GraphFilteringistheKeytoDesigningGraphConvolutionalNeuralNetworks
![Page 9: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/9.jpg)
9
GraphConvolutionalNetworksLearnaSequenceofGraphFilters
Chebyshev Approximation
![Page 10: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/10.jpg)
10
Problem:UsingGraphCNNforAutismSpectrumDisorderClassification
Neuropathologystudiesmapvariationsinbrainfunctionalitytoclinicalmeasures
Population Graph Design Signal ConstructionNon-imaging features(e.g.
gender/site)orcombinationofimagingandnon-imaging
features
Statisticsfrom imagingfeatures(e.g.fMRI)
![Page 11: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/11.jpg)
11
Fine-TuningGraphCNNsisHighlyChallengingduetotheirSensitivitytoGraphDesign
Population graph
G-CNN
G-CNN
G-CNN
Features
Features
Features
Consensus Class Label
Random Graph Ensemble
Predictive Modeling
Softmax Output
N-dimensional feature per node
Weadoptabootstrappingapproachthatperformsdropoutintheinputlayerofthenetwork
![Page 12: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/12.jpg)
12
Results:AutismBrainImagingDataExchangeInitiative§ 872patientsfrom20differentsites
§ SignalConstruction:Upper-Triangularpartofcovariancematrix+PCA
BaselinesLinear SVM 64.71Kernel SVM 65.72
Ensemble G-CNN(Gender, Site) 67.7
Linear Kernel + (Gender, Site) 67.7
Graph Kernel + (Gender, Site) 68.28
BootstrappedGraphCNNsimprovethepredictionaccuracyonthechallengingABIDEdataset
![Page 13: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/13.jpg)
13
§ Samesignaldefinedondifferentgraphscanleadtocompletelydifferentresults.
Problem:NeighborhoodGraphConstructionisSensitivetoSampleDensityandNoise
![Page 14: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/14.jpg)
14
Idea1:GraphAuto-EncoderscanRecovertheOptimalEmbeddingsforData
§ Auto-encodersaimtominimizethereconstructionloss
§ The“reconstructionnature”ofspectralclusteringcanbeviewedasanauto-encoder
§ UsingL2loss– HiddenfeatureswillconvergetothesmallesteigenvectorsoftheLaplacianmatrix
Laplacian Matrix
![Page 15: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/15.jpg)
15
Idea2:SpectralEmbeddingwithQuantileLossEnablesRobustSparsificationofGraphs
PiecewiseLinearQuadraticLosstypicallyusedtoexploreheterogeneousdatasets
Low-rank decomposition of the similarity matrix
![Page 16: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/16.jpg)
16
Idea2:SpectralEmbeddingwithQuantileLossEnablesRobustSparsificationofGraphs
PiecewiseLinearQuadraticLosstypicallyusedtoexploreheterogeneousdatasets
Low-rank decomposition of the similarity matrix
Whenthequantileparameterislarge:§ Positiveresidualswillbepenalizedlesser,i.e.theapproximationwill
underestimatethesimilarityfunction
Whenthequantileparameterissmall:§ Returnsadensegraphwithoverestimatededgestrengths
![Page 17: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/17.jpg)
17
ConditionalQuantilesoftheSimilarityFunctionRevealtheImportanceofEdges
Number of edges at 0.1 quantile
Highest quantile at which edge persists
Minimum probability for any edge
Rate of Decay Edge Probabilities
![Page 18: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/18.jpg)
18
§ Givenafewlabeledexamplesandaneighborhoodgraph,propagatelabelstoallsamples– Greedyrandomwalk
UsingtheLocally-ScaledGraphsinSemiSupervisedLabelPropagation
ExtremeCase– 1LabeledNode
![Page 19: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/19.jpg)
19
§ Graphsignalprocessingprovidesaconvenientframeworkforawidevarietyofdataanalysisproblems.
§ Canbehighlyeffectiveinfusingimagingandnon-imagingfeaturesinaneuralnetworksetting
§ Generalizeddeepnetworksfordataonnon-uniformgrids– thenewkidontheblock
§ Useofappropriatelossfunctionscanleadtorobustgraphconstructions
§ Appliestoproblemsinsupervised,unsupervised,semi-supervised,transferlearning…
Conclusions
![Page 20: Graph Signal Processing using Deep Neural Networks · 2017. 8. 31. · 19 §Graph signal processing provides a convenient framework for a wide variety of data analysis problems. §Can](https://reader036.fdocuments.us/reader036/viewer/2022071108/5fe3141874a5647e636c52c0/html5/thumbnails/20.jpg)
ContactJayaramanJ.ThiagarajanCenterforAppliedScientificComputingLawrenceLivermoreNationalLaboratoryEmail:[email protected]
Questions?