Explainable (Deep) Learning and Simulation approaches Torsten - - PowerPoint PPT Presentation

▶
explainable deep learning and simulation approaches
SMART_READER_LITE
LIVE PREVIEW

Explainable (Deep) Learning and Simulation approaches Torsten - - PowerPoint PPT Presentation

Jun 21, 2023 33 likes •1.04k views

Explainable (Deep) Learning and Simulation approaches Torsten Mller Visualization and Data Analysis University of Vienna Explainable (Deep) Learning and Simulation approaches Torsten Mller Visualization and Data Analysis

slide-1
SLIDE 1

Explainable 
 (Deep) Learning and Simulation approaches

Torsten Möller Visualization and Data Analysis University of Vienna

slide-2
SLIDE 2

Explainable 
 (Deep) Learning and Simulation approaches

Torsten Möller Visualization and Data Analysis University of Vienna

slide-3
SLIDE 3

Torsten Möller Summerschool, Sep 2018

Outline today

  • Why explainable?
  • the promise of data science
  • societal factors
  • How?
  • a process model for simulations
  • (machine) learning process

3

slide-4
SLIDE 4

Torsten Möller Summerschool, Sep 2018

Visual Data Science

4

slide-5
SLIDE 5

Torsten Möller Summerschool, Sep 2018

Overview

  • Data Science is all about modelling
  • The three types of modelling
  • Computational modelling
  • Statistical modelling
  • Empirical modelling
  • Challenges of Visual Data Science
  • Conclusions

5

slide-6
SLIDE 6

Torsten Möller Summerschool, Sep 2018

What is data science?

  • Dhar 2013: “Data Science is the study of

the generalizable extraction of knowledge from data.”

6 Vasant Dhar, “Data Science and Prediction”, (2013)

slide-7
SLIDE 7

Torsten Möller Summerschool, Sep 2018

What is data science?

  • Dhar 2013: “Data Science is the study of

the generalizable extraction of knowledge from data.”

  • Data Science is the study of exploration,

abstraction, and communication of complex systems through models from data.

7

slide-8
SLIDE 8

Torsten Möller Summerschool, Sep 2018

What is data science?

  • Dhar 2013: “Data Science is the study of

the generalizable extraction of knowledge from data.”

  • Data Science is the study of exploration,

abstraction, and communication of complex systems through models from data.

8

slide-9
SLIDE 9

Torsten Möller Summerschool, Sep 2018

What is data science?

  • Dhar 2013: “Data Science is the study of

the generalizable extraction of knowledge from data.”

  • Data Science is the study of exploration,

abstraction, and communication of complex systems through models from data.

9

slide-10
SLIDE 10

Torsten Möller Summerschool, Sep 2018

What is data science?

  • Dhar 2013: “Data Science is the study of

the generalizable extraction of knowledge from data.”

  • Data Science is the study of exploration,

abstraction, and communication of complex systems through models from data.

10

slide-11
SLIDE 11

Torsten Möller Summerschool, Sep 2018

Data Science

  • Jeff Leek: “The key word in ‘Data Science’ is

not Data, it is Science” “The issue is that the hype around big data/ data science will flame out (it already is) if data science is only about "data" and not about "science". The long term impact of data science will be measured by the scientific questions we can answer with the data.”

11

http://simplystatistics.org/2013/12/12/the-key-word-in-data-science-is-not-data-it-is-science/

slide-12
SLIDE 12

Torsten Möller Summerschool, Sep 2018

Overview

  • Data Science is all about modelling
  • The three types of modelling
  • Computational modelling
  • Statistical modelling
  • Empirical modelling
  • Challenges of Visual Data Science
  • Conclusions

12

slide-13
SLIDE 13

Torsten Möller Summerschool, Sep 2018

Scientific Method

13

Real world A model Hypothesis Observation Validation

after Hans Christian Ørsted, "First Introduction to General Physics" (1811)

slide-14
SLIDE 14

Torsten Möller Summerschool, Sep 2018

Validation ➙ Prediction

14

Real world A model Hypothesis Observation Validation Prediction

slide-15
SLIDE 15

Torsten Möller Summerschool, Sep 2018

4 Paradigms of Science

  • empirical: observe, then derive

15

Real world A model Hypothesis Observation Prediction

Jim Gray, “eScience -- A Transformed Scientific Method”, (2007)

https://en.wikipedia.org/wiki/File:Jim_Gray_portrait,_1999.jpg 1944-2007

slide-16
SLIDE 16

Torsten Möller Summerschool, Sep 2018

4 Paradigms of Science

  • empirical: observe, then derive
  • predictive: derive, then observe

16

Real world A model Hypothesis Observation

Jim Gray, “eScience -- A Transformed Scientific Method”, (2007)

https://en.wikipedia.org/wiki/File:Jim_Gray_portrait,_1999.jpg 1944-2007

Prediction

slide-17
SLIDE 17

Torsten Möller Summerschool, Sep 2018

4 Paradigms of Science

  • empirical: observe, then derive
  • predictive: derive, then observe
  • computational: simulate

17

computational model Real world A model Hypothesis Observation

Jim Gray, “eScience -- A Transformed Scientific Method”, (2007)

https://en.wikipedia.org/wiki/File:Jim_Gray_portrait,_1999.jpg 1944-2007

Prediction

slide-18
SLIDE 18

Torsten Möller Summerschool, Sep 2018

4 Paradigms of Science

  • empirical: observe, then derive
  • predictive: derive, then observe
  • computational: simulate
  • data-driven: measure

18

Real world Hypothesis Data

Jim Gray, “eScience -- A Transformed Scientific Method”, (2007)

https://en.wikipedia.org/wiki/File:Jim_Gray_portrait,_1999.jpg 1944-2007

computational model Prediction

slide-19
SLIDE 19

Torsten Möller Summerschool, Sep 2018

Three types of modelling

  • computational: the simulation of

discretized mathematical models (computational science)

  • statistical: data-driven — extracting

statistical models from data

  • empirical: simple, often linear models

19

slide-20
SLIDE 20

Torsten Möller Summerschool, Sep 2018

Computational Modelling

  • (almost) every discipline has these models
  • Examples:
  • Navier-Stokes, Maxwell, etc.
  • Population Dynamics
  • computational science: experimentation

through simulation of discretized models

20

slide-21
SLIDE 21

Torsten Möller Summerschool, Sep 2018

Vismon: Fisheries Science

21 Booshehrian, “Vismon: Facilitating Risk Assessment and Decision Making In Fisheries Management”, (2012)

slide-22
SLIDE 22

Torsten Möller Summerschool, Sep 2018 22

[Potter et al. 2009] [Bruckner & Möller 2010] [Bergner et al. 2013] [Coffey et al. 2013]

slide-23
SLIDE 23

Torsten Möller Summerschool, Sep 2018

Statistical Modeling

  • “Mainstream” understanding of Data Science
  • Classical (machine learning) approaches:
  • Clustering
  • Classification
  • Regression
  • (dimensionality reduction, outlier

detection, etc)

23

slide-24
SLIDE 24

Torsten Möller Summerschool, Sep 2018 24

Dim reduction — [Ingram et al. 2010] Regression — [Mühlbacher & Piringer 2013] Classification — [Linhardt et al. 2018?] Clustering — [Sedlmair et al. 2018?]

slide-25
SLIDE 25

Torsten Möller Summerschool, Sep 2018

Empirical Modeling

  • often no explicit modelling or only

simple models, e.g.

  • linear models
  • weighted averages etc.
  • examples: spreadsheets, rankings

25

slide-26
SLIDE 26

Torsten Möller Summerschool, Sep 2018

LineUp: Gratzl et al. 2013

26

slide-27
SLIDE 27

Torsten Möller Summerschool, Sep 2018

LineUp: Gratzl et al. 2013

27

slide-28
SLIDE 28

Torsten Möller Summerschool, Sep 2018 28

World Lines — [Waser et al. 2010] ValueCharts — [Carenini et al. 2004] Design Galleries — [Marks et al. 1997]

slide-29
SLIDE 29

Torsten Möller Summerschool, Sep 2018

Not just Labcoat Science

  • valid for business,

engineering, public policy

  • general data

analysis approach

29

Real world A model Hypothesis Data Prediction

slide-30
SLIDE 30

Torsten Möller Summerschool, Sep 2018

Overview

  • Data Science is all about modelling
  • The three types of modelling
  • Computational modelling
  • Statistical modelling
  • Empirical modelling
  • Challenges of Visual Data Science
  • Conclusions

30

slide-31
SLIDE 31

Torsten Möller Summerschool, Sep 2018

What is visual data science?

  • Visual Data Science is helping users

explore, abstract, and communicate complex systems through models from data.

31

slide-32
SLIDE 32

Torsten Möller Summerschool, Sep 2018

Acting upon models

32

Decisions Models
 (predictions) Data

slide-33
SLIDE 33

Torsten Möller Summerschool, Sep 2018

Building vs. Using

  • building models
  • computational

experts

  • bioinformaticians

33

Decisions Models
 (predictions) Data

  • using models
  • decision makers
  • domain experts
  • biologists
slide-34
SLIDE 34

Torsten Möller Summerschool, Sep 2018

Building vs. Using

  • building models
  • validation
  • uncertainty

34

Decisions Models
 (predictions) Data

  • using models
  • trust
  • tradeoffs + risks
slide-35
SLIDE 35

Torsten Möller Summerschool, Sep 2018

A modern microscope

  • making difficult algorithmic solutions

accessible to a broad audience: enable model users to become model builders

35

Decisions Models
 (predictions) Data

slide-36
SLIDE 36

Torsten Möller Summerschool, Sep 2018

What is a model?

  • has input parameters
  • creates outputs
  • it’s really “just” an algorithm

36

Direct Output Input Model

Sedlmair, “Visual Parameter Space Analysis: A Conceptual Framework”, (2014)

slide-37
SLIDE 37

Torsten Möller Summerschool, Sep 2018

What is a model?

  • paradigm shift:
  • from single input/output exploration to

input ranges and ensemble outputs

37

Model

Setting A Setting B Setting C Setting D Setting E

Sedlmair, “Visual Parameter Space Analysis: A Conceptual Framework”, (2014)

slide-38
SLIDE 38

Torsten Möller Summerschool, Sep 2018

Supporting the user

  • hypothesis creation
  • uncertainty / risk analysis
  • sensitivity analysis / model uncertainty
  • decision making / sense making

38

Decisions Models
 (predictions) Data

slide-39
SLIDE 39

Torsten Möller Summerschool, Sep 2018

Conclusions

39

slide-40
SLIDE 40

Torsten Möller Summerschool, Sep 2018

What is visual data science?

  • Visual Data Science is helping users

explore, abstract, and communicate complex systems through models from data.

40

slide-41
SLIDE 41

Torsten Möller Summerschool, Sep 2018

Three types of modelling

  • computational
  • statistical
  • empirical

41

slide-42
SLIDE 42

Torsten Möller Summerschool, Sep 2018

A modern microscope

  • making difficult algorithmic solutions

accessible to a broad audience: enable model users to become model builders

42

Decisions Models
 (predictions) Data

slide-43
SLIDE 43

Torsten Möller Summerschool, Sep 2018

Modern microscope

43

k-means kNN scatterplot

= =

DBScan SVM Isomap scatterplot + ? +

Visual Data Science

Making modelling techniques accessible to a broad set of users without requiring a PhD in Stats/ ML.

slide-44
SLIDE 44

Torsten Möller Summerschool, Sep 2018

Why?: Societal factors

44

slide-45
SLIDE 45

Torsten Möller Summerschool, Sep 2018

Ethics

  • cars make decisions on who to run over and

who not

  • who should the company hire?
  • which update from which friend should you be

shown?

  • which convict is more likely to re-offend?
  • which news item / movie should we

recommend to people?

45 https://www.ted.com/talks/zeynep_tufekci_machine_intelligence_makes_human_morals_more_important#t-157020

slide-46
SLIDE 46

Torsten Möller Summerschool, Sep 2018

Laws

  • EU’s General Data Protection Regulation:
  • incl Article 22: Automated individual decision-making, including profiling
  • prohibits any “decision based solely on automated processing, including

profiling” which “significantly affects” a data subject

  • Discrimination: Paragraph 71 of the recitals (the preamble to the GDPR,

which explains the rationale behind it but is not itself law) explicitly requires data controllers to “implement appropriate technical and

  • rganizational measures” that “prevents, inter alia, discriminatory effects”
  • n the basis of processing sensitive data
  • Right to explanation: Articles 13 and 14 state that, when profiling takes

place, a data subject has the right to “meaningful information about the logic involved.”

46 Goodman, B. & Flaxman, S. European Union regulations on algorithmic decision-making and a “right to explanation” AI Magazine, 2017

slide-47
SLIDE 47

Torsten Möller Summerschool, Sep 2018

Outline today

  • Why explainable?
  • the promise of data science
  • extrinsic factors
  • How?
  • a process model for simulations
  • (machine) learning environments

47

slide-48
SLIDE 48

Torsten Möller Summerschool, Sep 2018

How?

48

From Philip Grohs

slide-49
SLIDE 49

Torsten Möller Summerschool, Sep 2018

How?

49

Alex Schindler

slide-50
SLIDE 50

Torsten Möller Summerschool, Sep 2018

How?

50

Alex Schindler

slide-51
SLIDE 51

Torsten Möller Summerschool, Sep 2018

How — our approach

https://youtu.be/5d71xhEbjDg

51

slide-52
SLIDE 52

Torsten Möller Summerschool, Sep 2018

FluidExplorer Fluid animation

52

slide-53
SLIDE 53

Torsten Möller Summerschool, Sep 2018

Special effects

  • Fluid simulation is

heavily used in the motion picture industry

  • Most common

animation packages include solvers or offer add-ons

  • Problem: Difficult to

control for visual

53

slide-54
SLIDE 54

Torsten Möller Summerschool, Sep 2018

Special effects (2)

  • Tens of parameters
  • Hard to predict results
  • Time-consuming trial &

error

54

Autodesk Maya 2010

slide-55
SLIDE 55

Torsten Möller Summerschool, Sep 2018

Overview

55

slide-56
SLIDE 56

Torsten Möller Summerschool, Sep 2018

Visualization

  • Body Level One
  • Body Level Two
  • Body Level Three
  • Body Level Four
  • Body Level Five

56

slide-57
SLIDE 57

Torsten Möller Summerschool, Sep 2018

Abstraction:
 (visual) Parameter space exploration (vPSA)

57

slide-58
SLIDE 58

Torsten Möller Summerschool, Sep 2018

Other tools

58

slide-59
SLIDE 59

Torsten Möller Summerschool, Sep 2018

Much recent attention in vPSA

  • Image segmentation [Torsney Weir et al. 2011]
  • Weather forecast [Potter et al. 2009]
  • Disaster simulation [Waser et al. 2010]
  • many more …

59

[Pretorius et al. 2011] [Waser et al. 2010] [Coffey et al. 2013] [Potter et al. 2009] [Torsney-Weir et al. 2011] [Piringer et al. 2010] [Bruckner & Möller 2010] [Amirkhanov et al. 2010] [Bergner et al. 2013]

…etc.

slide-60
SLIDE 60

Torsten Möller Summerschool, Sep 2018

Much recent attention in vPSA

  • Image segmentation [Torsney Weir et al. 2011]
  • Weather forecast [Potter et al. 2009]
  • Disaster simulation [Waser et al. 2010]
  • many more …

60

[Pretorius et al. 2011] [Waser et al. 2010] [Coffey et al. 2013] [Potter et al. 2009] [Torsney-Weir et al. 2011] [Piringer et al. 2010] [Bruckner & Möller 2010] [Amirkhanov et al. 2010] [Bergner et al. 2013]

…etc.

slide-61
SLIDE 61

Torsten Möller Summerschool, Sep 2018

Much recent attention in vPSA

  • Image segmentation [Torsney Weir et al. 2011]
  • Weather forecast [Potter et al. 2009]
  • Disaster simulation [Waser et al. 2010]
  • many more …

61

[Pretorius et al. 2011] [Waser et al. 2010] [Coffey et al. 2013] [Potter et al. 2009] [Torsney-Weir et al. 2011] [Piringer et al. 2010] [Bruckner & Möller 2010] [Amirkhanov et al. 2010] [Bergner et al. 2013]

…etc.

slide-62
SLIDE 62

Torsten Möller Summerschool, Sep 2018

Much recent attention in vPSA

  • Image segmentation [Torsney Weir et al. 2011]
  • Weather forecast [Potter et al. 2009]
  • Disaster simulation [Waser et al. 2010]
  • many more …

62

[Pretorius et al. 2011] [Waser et al. 2010] [Coffey et al. 2013] [Potter et al. 2009] [Torsney-Weir et al. 2011] [Piringer et al. 2010] [Bruckner & Möller 2010] [Amirkhanov et al. 2010] [Bergner et al. 2013]

…etc.

slide-63
SLIDE 63

Torsten Möller Summerschool, Sep 2018

Much recent attention in vPSA

  • comprehensive study of 21 different tools

63

[Pretorius et al. 2011] [Waser et al. 2010] [Coffey et al. 2013] [Potter et al. 2009] [Torsney-Weir et al. 2011] [Piringer et al. 2010] [Bruckner & Möller 2010] [Amirkhanov et al. 2010] [Bergner et al. 2013]

…etc.

slide-64
SLIDE 64

Torsten Möller Summerschool, Sep 2018

Data Flow Model

64

slide-65
SLIDE 65

Torsten Möller Summerschool, Sep 2018

Build an estimator

65

slide-66
SLIDE 66

Torsten Möller Summerschool, Sep 2018

Model Input Output

66

slide-67
SLIDE 67

Torsten Möller Summerschool, Sep 2018

Model Input Output

Model

  • simulation model, prediction model, …
  • … but also algorithm
  • stochastic, deterministic
  • usually black box (to us as Vis researchers)

67

slide-68
SLIDE 68

Torsten Möller Summerschool, Sep 2018

Model Input Output

Inputs

  • well chosen by the scientist, i.e. people care about

their inputs

  • normally continuous (quantitative data)
  • need to sample the space
  • categorical data common too (e.g. use of a

different algorithm)

68

slide-69
SLIDE 69

Torsten Möller Summerschool, Sep 2018 69

Model Input Output

Outputs

  • typically complex objects, e.g.
  • 2D, 3D images (Tuner)
  • animations (FluidExplorer)
  • performance graphs (fuel cells)
  • hard to evaluate / compare many complex outputs
slide-70
SLIDE 70

Torsten Möller Summerschool, Sep 2018 70

Derive Derived Outputs Model Input Output

Derive

  • one-dimensional (“goodness”) rating: d(O1)
  • two-dimensional comparison: d(O1, O2)
  • objective measures can be
  • exact (reliable)
  • approximate - about right, but not 100% precise
  • unknown (active learning)
slide-71
SLIDE 71

Torsten Möller Summerschool, Sep 2018

Complex objects (in 18/21 papers)

71

Model

1.0 2.1 3.7

?

1.0 2.1 3.7 6.3 3.3 5.2 2.2 2.1 2.0 1.1 5.6 7.8 … … … Input Parameters … … Outputs

[Torsney-Weir et al. 2011]

slide-72
SLIDE 72

Torsten Möller Summerschool, Sep 2018

Derive objective measures

72

Model Derive

1.0 2.1 3.7

7.1

slide-73
SLIDE 73

Torsten Möller Summerschool, Sep 2018

Surrogate models

73

?

Model Derive

?

1.5 2.5 3.5

expensive!

slide-74
SLIDE 74

Torsten Möller Summerschool, Sep 2018

Surrogate models

74

Surrogate Model Model Derive

1.5 2.5 3.5 1.5 2.5 3.5

slide-75
SLIDE 75

Torsten Möller Summerschool, Sep 2018

Data flow model

75

Direct Output Derived Output Predicted Output Input Model Derive Surrogate Model

slide-76
SLIDE 76

Torsten Möller Summerschool, Sep 2018

Navigation Strategies

76

slide-77
SLIDE 77

Torsten Möller Summerschool, Sep 2018

Navigation strategies

  • Trial and error (traditional approach)

77

slide-78
SLIDE 78

Torsten Möller Summerschool, Sep 2018

Navigation strategies

  • Trial and error (traditional approach)
  • Local —> global tweaking

78

Design by Dragging [Coffey et al., SciVis 2013]

slide-79
SLIDE 79

Torsten Möller Summerschool, Sep 2018

Navigation strategies

  • Trial and error (traditional approach)
  • Local —> global tweaking
  • Global —> local exploration
  • FluidExplorer, Vismon, Tuner
  • many others: Paramorama [Pretorius et al., InfoVis 2011]

79

slide-80
SLIDE 80

Torsten Möller Summerschool, Sep 2018

Navigation strategies

  • Trial and error (traditional approach)
  • Local —> global tweaking
  • Global —> local exploration

80

  • Steering
  • simulation steering: e.g.

real-time simulators

  • computational steering:

e.g. change the grid size, stop if no insight

World Lines [Waser et al., Vis 2010]

slide-81
SLIDE 81

Torsten Möller Summerschool, Sep 2018

Analysis Tasks

81

slide-82
SLIDE 82

Torsten Möller Summerschool, Sep 2018

Analysis tasks

  • Optimization
  • Partitioning
  • Fitting
  • Outliers
  • Uncertainty
  • Sensitivity

82

slide-83
SLIDE 83

Torsten Möller Summerschool, Sep 2018

Analysis tasks

1 1

  • Optimization
  • Partitioning
  • Fitting
  • Outliers
  • Uncertainty
  • Sensitivity

83

1 1

Model

1 2 3 3 4 5

Find the best parameter combination given some

  • bjectives.

in 19/21 papers

slide-84
SLIDE 84

Torsten Möller Summerschool, Sep 2018

Analysis tasks

  • Optimization
  • Partitioning aka clustering
  • Fitting
  • Outliers
  • Uncertainty
  • Sensitivity

84

How many different types of model behaviors are possible?

Model

1 2 3 1 1 3 4 5 1 2 1 3 3 2 1 3 3 5 1 4 3 4 3 5

in 6/21 papers

slide-85
SLIDE 85

Torsten Möller Summerschool, Sep 2018

Analysis tasks

  • Optimization
  • Partitioning
  • Fitting aka regression analysis
  • Outliers
  • Uncertainty
  • Sensitivity

85

Where in the input parameter space would actual measured data

  • ccur?

Model Derive

ground truth

in 9/21 papers

slide-86
SLIDE 86

Torsten Möller Summerschool, Sep 2018

Analysis tasks

  • Optimization
  • Partitioning
  • Fitting
  • Outliers
  • Uncertainty
  • Sensitivity

86

What outputs are special?

Model

in 9/21 papers

slide-87
SLIDE 87

Torsten Möller Summerschool, Sep 2018

Analysis tasks

  • Optimization
  • Partitioning
  • Fitting
  • Outliers
  • Uncertainty
  • Sensitivity

87

How reliable is the output?

Model

  • model vs. reality
  • non-deterministic 


model

  • model vs. surrogate

in 7/21 papers

slide-88
SLIDE 88

Torsten Möller Summerschool, Sep 2018

Analysis tasks

  • Optimization
  • Partitioning
  • Fitting
  • Outliers
  • Uncertainty
  • Sensitivity

88

What ranges/variations of

  • utputs to expect with changes
  • f input?

Model

in 14/21 papers

slide-89
SLIDE 89

Torsten Möller Summerschool, Sep 2018

The (machine) learning process

89

slide-90
SLIDE 90

Torsten Möller Summerschool, Sep 2018

types of learning

  • regression
  • classification (supervised)
  • clustering (unsupervised)
  • (dimensionality reduction)
  • (outlier detection)

90

slide-91
SLIDE 91

Torsten Möller Summerschool, Sep 2018

techniques of learning

  • Neural Networks (plus Deep-NN)
  • Kernel methods (SVM)
  • Graphical models
  • Ensemble methods
  • …

91

The world of ML algorithms is not as well organized in terms of strategies as it is with simulation environments. This is work in progress.

slide-92
SLIDE 92

Torsten Möller Summerschool, Sep 2018

A small selection:

  • confusion matrixes for classification
  • deep neural nets
  • understand / diagnose / refine
  • Explainers
  • LIME

92

slide-93
SLIDE 93

Torsten Möller Summerschool, Sep 2018

Confusion matrix

  • Google’s Facet:
  • http://gifctrl.com/?g=https://

3.bp.blogspot.com/-T0dTxdse9Ow/ WWz0u431RpI/AAAAAAAAB5M/ rBvToJjx1L0FVVpXkgNOAwzXASyZC_JWw CLcBGAs/s1600/image4.gif

  • EuroVis keynote, 2017 — https://

www.youtube.com/watch?v=E70lG9-HGEM

93

slide-94
SLIDE 94

Torsten Möller Summerschool, Sep 2018

Squares

94 Ren et al., 2017. Squares: Supporting inter- active performance analysis for multiclass classifiers. IEEE TVCG 23 (1), 61–70.

slide-95
SLIDE 95

Torsten Möller Summerschool, Sep 2018

Deep NN’s: 
 Neurons — point based

95 Rauber, et al., 2017. Visualizing the hidden activity of artificial neural networks. IEEE TVCG 23 (1), 101–110

slide-96
SLIDE 96

Torsten Möller Summerschool, Sep 2018

Deep NN’s: 
 Neurons — network based

96 Tzeng, F.Y., Ma, K.L. 2005. Opening the black box - data driven visualization of neural networks. In: IEEE Visualization

slide-97
SLIDE 97

Torsten Möller Summerschool, Sep 2018

CNNVis

97

http://shixialiu.com/publications/cnnvis/demo/

slide-98
SLIDE 98

Torsten Möller Summerschool, Sep 2018

Conclusions

  • Why explainable?
  • improve algorithms
  • trust
  • bridge the model builder / model usage gap
  • ethics and law
  • How?
  • characterization of input-output relationships OR parameter tuning
  • understanding the behaviour of neurons in Deep NN
  • It is the “wild west” in terms of understanding machine learning

models!

98

slide-99
SLIDE 99

Torsten Möller Summerschool, Sep 2018

Acknowledgments

Tamara Munzner UBC Tom Torsney-Weir U of Vienna Maryam Booshehrian Muprime Tech Melanie Tory Tableau Steven Bergner SFU Stefan Bruckner U of Bergen

99

Stephen Ingram Coho Data Michael Sedlmair U of Vienna Harald Piringer VRVis Hamid Younesy
 SFU Lorenz Linhardt ETH Zurich Patrick Wolf Software Dev

slide-100
SLIDE 100

Torsten Möller Summerschool, Sep 2018

References

  • Visual Parameter Space Analysis: A Conceptual Framework. Michael Sedlmair, Christoph

Heinzl, Stefan Bruckner, Harald Piringer, Torsten Möller, IEEE Transactions on Visualization and Computer Graphics 20(12):2161-2170, 2014.

  • eScience -- A Transformed Scientific Method. Jim Gray, (2007), in “The Fourth Paradigm:

Data-Intensive Scientific Discovery”, 2009.

  • Google Facet, https://pair-code.github.io/facets/, Jul 2017.
  • Squares: Supporting Interactive Performance Analysis for Multiclass Classifiers, D. Ren,
  • S. Amershi, B. Lee, J. Suh and J. D. Williams, IEEE Transactions on Visualization and

Computer Graphics, vol. 23, no. 1, pp. 61-70, Jan. 2017.

  • Visualizing the Hidden Activity of Artificial Neural Networks, P. E. Rauber, S. G. Fadel,
  • A. X. Falcão and A. C. Telea, IEEE Transactions on Visualization and Computer Graphics,
  • vol. 23, no. 1, pp. 101-110, Jan. 2017.
  • Towards Better Analysis of Deep Convolutional Neural Networks. Mengchen Liu, Jiaxin

Shi, Zhen Li, Chongxuan Li, Jun Zhu, and Shixia Liu. IEEE Transactions on Visualization and Computer Graphics 23, 1 (January 2017), 91-100.

100

slide-101
SLIDE 101

Torsten Möller Summerschool, Sep 2018

Questions?

http://vda.cs.univie.ac.at torsten.moeller@univie.ac.at

101