hierarchical stochastic neighbor embedding
TRANSCRIPT
PowerPoint Presentation
Hierarchical Stochastic Neighbor EmbeddingNicola Pezzotti1, Thomas Hllt1, Boudewijn P.F. Lelieveldt2, Elmar Eisemann1,Anna Vilanova1Computer Graphics & Visualization, Delft University of Technology, Delft, The NetherlandsDivision of Image Processing, Leiden Medical Center, Leiden, The Netherlands
EuroVis 201618th EG/VGTC Conference on Visualization6-10 June 2016, Groningen, the Netherlands
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
1
Hierarchical organization of data
Image Collection
Nature
Man-made
Ships
Vehicles
2
EuroVis 2016
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //Images, they can be organized hierarchically based on the objects that they represent.And we can do that for different data
This kind of hierarchies arise when we
Visualizing relationships between data pointsParallel-Coordinate Plots do not scaleDimensionality Reduction (DR)3Embedding
High-DimensionalFeature Vectors
DimensionalityReduction
Dim-1Dim-2
DataFeatureExtraction
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //Non-linear Dimensionality ReductionData often lay on a non-linear manifold in the high-dimensional spaceWidely used on real-world dataComputationally intensive4
EuroVis 2016
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Non-Linear DR with Landmarks5
[Landmark-SNE, Landmark-ISOMAP][LSP, P-LSP, LAMP, LoCH, Pekalska]
Hybrid techniques
Non linear
Dim-1Dim-2Emb-Dim-1Emb-Dim-1
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Multiscale Dimensionality Reduction Non-linear DRLandmark basedHierachical exploration of the dataOverview-first & Details-on-DemandFilter & Drill-inProabilistic framework
Hierarchical Stochastic Neighbor Embedding6
Hierarchical SNEEmb-Dim-1
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //Algorithm
EuroVis 2016
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm8
SimilarityBased Embedding tSNE1
1: Van der Maaten et al. - Visualizing data using t-SNE - Journal of Machine Learning Research - 2008.Localized SimilaritiesLow memory footprint
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm9
Dim-1Dim-2
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm10
Dim-1Dim-2
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm11
Dim-1Dim-2
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm12
Dim-1Dim-2
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm13
Dim-1Dim-2
LowHighDistribution
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm14
Dim-1Dim-2
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm15
Dim-1Dim-2
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm16
Dim-1Dim-2
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm17
Dim-1Dim-2
66%33%Localized Area of InfluenceLow memory footprint
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm18
SimilarityBased Embedding tSNE
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm19
SimilarityBased Embedding tSNE
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Hierarchical SNE - Algorithm20
Random walksMore than 1k per ms
Hierarchical AnalysisTop-downLink between scale given by the area of influence
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Use Case 1Deep Learning
EuroVis 2016
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //Use case I: Deep Learning22
Feature vector 4096 DimensionsAre the images processed by AlexNet [1] organized hierarchically by the network?1: Krizhevsky et al. - ImageNet Classification with Deep Convolutional Neural Networks - Advances in neural information processing systems - 2012.
Label+Image
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //Use case I: Deep Learning
Test set
NatureMan-made100k Images92s23
EuroVis 2016
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Use case I: Deep Learning
Nature
VehiclesAppliancesShipsMan-made
24
EuroVis 2016
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Use case I: Deep Learning
AppliancesShipsVehicles
TrainsCarsBusesSport cars
25
EuroVis 2016
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Use Case 2Hyperspectral Images
EuroVis 2016
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //Use case II: Hyperspectral images27
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //Use case II: Hyperspectral images28
Pixels 1M Data points (1024x1024)Images 12 DimensionsClusters in the EmbeddingGroup of pixels that correspond to the same phenomenon
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //Use case II: Hyperspectral images29
Surface
Space
LowHighInfluence
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Use case II: Hyperspectral images30
Outer space
Corona
LowHighInfluence
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Use case II: Hyperspectral images31
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Use case II: Hyperspectral images32
LowHighInfluence
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //Conclusion
EuroVis 2016
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //Hierarchical Stochastic Neighbor EmbeddingNovel hierarchical analysis of non-linear dataOutperforms existing techniquesComputation timeSize of the data to be computedK-Nearest Neighbor PreservationStability of the embeddings
34Conclusions
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //
Questions?
This project is founded by STW through the V.An.P.I.Re project
EuroVis 2016
Pezzotti et al. // Hierarchical Stochastic Neighbor Embedding //