Approximate and User Steerable tSNE for Progressive Visual Analytics Nicola Pezzotti , Boudewijn P.F. Lelieveldt, Laurens van der Maaten, Thomas Höllt, Elmar Eisemann, Anna Vilanova
Jan 16, 2017
PowerPoint Presentation
Approximate and User Steerable tSNE for Progressive Visual AnalyticsNicola Pezzotti, Boudewijn P.F. Lelieveldt, Laurens van der Maaten,Thomas Hllt, Elmar Eisemann, Anna Vilanova
1
Non-Linear Dimensionality-Reduction
#
Non-linear dimensionality-reduction algorithmPreserves small neighborhoodsReveals global structures
Visualizing data using t-SNE - Van der Maaten & Hinton - 2008
t-Distributed Stochastic Neighbor Embedding
#3
3
tSNESimilaritiesComputation
Similarities Computation
Gradient descent minimizationSimilarities
tSNE as a Black Box
#4
PVA - tSNESimilaritiesComputation
Similarities Computation
Gradient descent minimizationSimilaritiesProgressive Visual Analytics: User-Driven Visual Exploration of In-Progress Analytics - Stolper et al. - 2014Opening the Black Box: Strategies for Increased User Involvement in Existing Algorithm Implementations - Muhlbacher et al. - 2014Progressive Analytics: A Computation Paradigm for Exploratory Data Analysis - Fekete & Primet - 2016
Visualization
Compute partial results
tSNE
Progressive Visual Analytics (PVA)
#5
Approximated Computations in PVA
6
Approximated - tSNE
Similarities Computation
Similarities
Visualization
Compute partial results
ApproximatedSimilarities
PVA - tSNEApproximated tSNE
#7
ApproximatedK-Nearest-Neighborhood [1]Precision: 50%
[1] Fast Approximate Nearest Neighbors with Automatic Algorithm Configuration - Muja et al. - 2009K-Nearest-NeighborhoodApproximated similarities computation
#8
Approximated - tSNE
Similarities Computation
Similarities
Visualization
Compute partial results
ApproximatedSimilarities
Approx.Refinement
Exact Refinement
Approximated tSNE
#9
tSNETime: 3191.8 sA-tSNE Precision: 35%Time: 30.1 sSpeed up: 100xPrecision 35% ?
#10
Approximated similarities computation
#11
Density-based visualizationSupports brushing & linking
Approximation is visualized and removed if requested3 StrategiesLocal minima avoidance
Steerability & Approximation visualization
A-tSNE Precision: 5%Preprocessing: 12 s
#12
Case Study I : Gene Expression in the Mouse Brain
13
Case Study I : Gene expression
SagittalAxial3D VolumeCoronal61164 data points (Voxels) 4345 dimensions (Gene expression)
#14
Case Study I : Gene expressionA-tSNE 50 seconds tSNE 3 hours and 50 minutesSpeed up: 250x
#Case Study II : High-dimensional data streams
16
Case Study II : High-dimensional data streamsChest - Ankle - Wrist52 Dimensions every 100 ms
Image courtesy of www.activ8all.com
#17
[1] Hierarchical Stochastic Neighbor Embedding - Pezzotti et al. - 2016ConclusionsApproximation in Progressive Visual AnalyticsApproximated-tSNEData manipulationRefinement
Scalability issues of the gradient descentHierarchical SNE [1]
#18
Thank you for your attention!A-tSNEPrecision: 35%tSNEA-tSNEPrecision: 5%Similarities computation time: 12 sSimilarities computation time: 29 sPrecomp. 3195 sSpeed 4x29 s12 s
#19
#
#
#
#
#
#
#