Visualisation as a mean to tackle some ethical issues ... · Visualisation as a mean to tackle some...
Transcript of Visualisation as a mean to tackle some ethical issues ... · Visualisation as a mean to tackle some...
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Visualisation as a mean to tackle some ethical issues raised by Machine Learning
Dr Ir Benoît Otjacques
Head of Environmental Informatics Unit
Luxembourg Institute of Science and Technology
Invited Talk at Frankfurt Big Data Lab, Goethe University
16th April, 2019
16th April 2019Dr Ir Benoît Otjacques, [email protected]
After some issues in the past…
16th April 2019Dr Ir Benoît Otjacques, [email protected]
AI is now governed, safe and under control
https://www.microsoft.com/en-us/ai/our-approach-to-ai, 12th April, 2019Published 8th April 2019
https://ai.google/principles/
12th April, 2019
16th April 2019Dr Ir Benoît Otjacques, [email protected]
AI is now governed, safe and under control
IEEE Ethicall Aligned Design: 8 General Principles
1. Human Rights
A/IS shall be created and operated to respect, promote, and protect internationally
recognized human rights
2. Well-being
A/IS creators shall adopt increased human well-being as a primary success criterion for
development
3. Data Agency
A/IS creators shall empower individuals with the ability to access and securely share their
data, to maintain people’s capacity to have control over their identity
4. Effectiveness
A/IS creators and operators shall provide evidence of the effectiveness and fitness for
purpose of A/IS
A/IS: Autonomous and Intelligent Systems
16th April 2019Dr Ir Benoît Otjacques, [email protected]
AI is now governed, safe and under control
IEEE Ethicall Aligned Design: 8 General Principles
5. Transparency
The basis of a particular A/IS decision should always be discoverable
6. Accountability
A/IS shall be created and operated to provide an unambiguous rationale for all decisions
made
7. Awareness of Misuse
A/IS creators shall guard against all potential misuses and risks of A/IS in operation
8. Competence
A/IS creators shall specify and operators shall adhere to the knowledge and skill required
for safe and effective operation
A/IS: Autonomous and Intelligent Systems
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Outline
Machine
LearningVisualisation
Use of ML to design Visualisations?
Use of Visualisations to better use ML?
Ethics? Ethics?
Ethics?
Ethics?
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Outline
• Concepts
• Ethical issues in ML
• Ethical issues in Visualisation
• ML-supported visualisation
• Visu-supported ML
• Visu + ML
• Conclusion
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
Artificial Intelligence
“Artificial intelligence (AI) refers to systems that display intelligent
behaviour by analysing their environment and taking actions –
with some degree of autonomy – to achieve specific goals.
AI-based systems can be purely software-based, acting in the
virtual world (e.g. voice assistants, image analysis software,
search engines, speech and face recognition systems) or AI can
be embedded in hardware devices (e.g. advanced robots,
autonomous cars, drones or Internet of Things applications).”
Source: Communication from the Commission to the European Parliament, the European Council, the Council, the European Economic and Social
Committee and the Committee of the Regions on Artificial Intelligence for Europe, Brussels, 25.4.2018 COM(2018) 237 final.
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
Artificial Intelligence as a Scientific Discipline
AI
Machine Learning
Deep
Learning
Reinforc.
Learning
… Reasoning
Search /
Optimisation
Planning /
Scheduling
Knowledge Repr.
and Reasoning…
Robotics
Source: A definition of AI: Main capabilities and scientific disciplines
High-Level Expert Group on Artificial Intelligence, published 8 April 2019
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
Machine Learning
Machine Learning is the field of study that gives computers the
ability to learn without being explicitly programmed
(Arthur Samuel, 1959)
A computer program is said to learn from experience E with
respect to some task T and some performance measure P, if
its performance on T, as measured by P, improves with
experience E. (Tom Mitchell, 1997)
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
Information
Visualisation
Visual
Perception
Computer
Vision
Computer
GraphicsInfographics
Visual
Analytics
Scientific
Visualisation
Informative
Art
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
Visual Perception
Visual Perception is the process of acquiring knowledge about
environmental objects and events by extracting information
from the light they reflect or emit.(Stephen E. Palmer, Vision Science, MIT Press, 1999)
Kolb, American Scientist, 2003
Retina
Nakai et al., Clinical Neurophysiology, 2018
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
Computer Graphics
Computer Graphics: 3D Image Analysis and Synthesis that takes into
account the whole image processing pipeline from scene acquisition to
scene reconstruction to scene editing to scene rendering. We also take into
account human perception on all levels of the pipeline, and we exploit the
abundance of digital visual data to extract powerful priors that can assist us
in the various tasks. (Max Plank Institute Informatik, 2017)
Moloney et al., SIGGRAPH 2017 – Star Wars
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
Computer Vision
Computer vision is concerned with the automatic extraction,
analysis and understanding of useful information from a single
image or a sequence of images. (British Machine Vision Association and Society for Pattern Recognition, 2018)
Nuske et al., IEEE/RSJ 2011 Cordts et al., TPAMI 39(7), 2017
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
Scientific Visualisation (SciVis)
In SciVis, the graphical models are typically constructed from
measured or simulated data representing objects or concepts
associated with phenomena from the physical world. (Ferreira and Levkowitz, TVCG, 9(3), 2003)
Wald et al, TCVG, 23(1), 2017
Engineering
Klein et al., TCVG, 24(1), 2018
Cell Biology
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
Information Visualisation (Infovis)
Infovis is the use of computer-supported, interactive, visual
representations of abstract data in order to amplify cognition (Card, Mackinlay and Shneiderman, 1999)
Infovis is the communication of abstract data through the use of
interactive visual interfaces (Keim et al., 2006)
Ellimaps
Otjacques et al., 2007
Slice-and-dice treemaps
Johnson & Schneiderman, 1992
Voronoi treemaps
Balzer & Deussen, 2005
Weighted maps
Ghoniem et al., 2015
Hybrid treemap
Hahn & Döllner, 2017
To date, 161 distinct techniques have been identified to visualise trees with included shapes
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
Visual Analytics (VA)
Visual analytics is the science of analytical reasoning
facilitated by interactive visual interfaces (Illuminating the Path. The Research and Development Agenda for Visual Analytics, Ed.
JJ Thomas and K.A Cook, IEEE Editions, 2005)
Matkovic et al. TCVG 20(12), 2014
Common Rail Engine Design
Médoc et al.,VAST Challenge 2014
Crisis management
(InfoVis or SciVis) + Data Analytics
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
Infographics
Infographics are graphic visual representations of information, data or
knowledge intended to present information quickly and clearly.(Wikipedia, 2018)
visualcapitalist.com, 2017
Popular Science, June, 2017
Popular Science, July, 2017
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
Informative Art
Informative art is is a type of computer applications which
borrow their appearance from well-known artistic styles
to visualize dynamically updated information
(Redström et al. , 2000)
Aesthetic use of InfoVis or SciVis techniques
to convey an artistic message
Holmquist and Skog, 2003
Samanci and Snyder, Vis Arts, 2017
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
There are still other concepts…
Data Visualization
Visual Language
Graphic Design
Visual Communication
Visual Thinking
Knowledge Visualisation
…
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Concepts
…
Deep
Learning
Clustering
Regression
Descriptive
statisticsScatter plot
Space-
filling trees
Graph &
Networks
Parallel
coordinates
…
Histogram
Pie chart
Pixel-
based visu.
Dim.
Reduction How to combine visualisation techniques
and analytics methods
in a meaningful and ethical manner?
16th April 2019Dr Ir Benoît Otjacques, [email protected]
• Concepts
• Ethical issues in ML
• Ethical issues in Visualisation
• ML-supported visualisation
• Visu-supported ML
• Visu + ML
• Conclusion
Outline
16th April 2019Dr Ir Benoît Otjacques, [email protected]
ML & Ethics
Machine Learning
A computer program is said to learn from experience E with
respect to some task T and some performance measure P, if
its performance on T, as measured by P, improves with
experience E. (Tom Mitchell, 1997)
Ethics?
Ethics? Ethics?
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Machine Learning
A computer program is said to learn from experience E with respect
to some task T and some performance measure P, if its
performance on T, as measured by P, improves with experience E.
Ethics?
This white dog is a cat
with a confidence of 96%
Training phase using
biased dataset
ML & Ethics
16th April 2019Dr Ir Benoît Otjacques, [email protected]
When you see the pictures, is it so difficult to realize that the training dataset is biased?
ML & Ethics
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Machine Learning
A computer program is said to learn from experience E with respect
to some task T and some performance measure P, if its
performance on T, as measured by P, improves with experience E.
Ethics?ML & Ethics
Post-deployment on purpose adversarial attack
2 attacks
- Stop sign to be misclassified
as a Speed Limit sign in 100% of the testing conditions
- Right Turn sign to be misclassified as either a Stop or
Added Lane sign in 100% of the testing conditions
Source: Evtimov, Ivan; Eykholt Kevin; Fernandes Earlence; Kohno Tadayoshi; Li Bo; Prakash Atul; Rahmati Amir; and Dawn Song, Robust Physical-World Attacks
on Machine Learning Models, arXiv preprint 1707.08945, August 2017, accessed 15th August 2017.
Can we visualize which pixels in the image led to a misclassification?
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Source: Reuters, https://www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-secret-ai-
recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G, accessed 12th April, 2019
ML & Ethics
16th April 2019Dr Ir Benoît Otjacques, [email protected]
When you see the charts,
is it so difficult to realize that
the training dataset is biased?
Source: Reuters, https://www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-secret-ai-
recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G, accessed 12th April, 2019
ML & Ethics
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Machine Learning
A computer program is said to
learn from experience E with
respect to some task T and
some performance measure
P, if its performance on T, as
measured by P, improves with
experience E.
Ethics?
Source: https://www.fbo.gov/index.php?s=opportunity&mode=form&id=29a4aed941e7e87b7af89c46b165a091&tab=core&_cview=0,
accessed 12th April, 2019
What if T is operating an
autonomous weapon?
Feb., 2019ML & Ethics
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Machine Learning
A computer program is said to learn from experience E with respect
to some task T and some performance measure P, if its
performance on T, as measured by P, improves with experience E.
Ethics?
ML & Ethics
• Does P really measure an improvement of the ML model?
• Does P fairly reflect the real world?
• Does several Pi compete with each other wrt ethics?
Goodhart's law:
"When a measure becomes a target, it ceases to be a good measure."
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Machine Learning
A computer program is said to learn from experience E with respect
to some task T and some performance measure P, if its
performance on T, as measured by P, improves with experience E.
Ethics?ML & Ethics
Anscombe’s seminal paper:
“Graphs in Statistical Analysis” (1973)*
4 data sets, each comprising 11 (x,y) pairs
All data sets yields the same standard stats output
• Mean of the x values = 9.0
• Mean of the y values = 7.5
• Equation of linear regression: y = 3 + 0.5 x
• Multiple R2 = 0.667
• …
case 1
x y
10 8.04
8 6.95
13 7.58
9 8.81
11 8.33
14 9.96
6 7.24
4 4.26
12 10.84
7 4.82
5 5.68
* The American Statistician, Vol. 27 (1), pp. 17-21
16th April 2019Dr Ir Benoît Otjacques, [email protected]
ML & Ethics
Seeing the graphics makes identical values for R2 less convincing
to adopt as a model the linear regression y = 3 + 0.5 x
* The American Statistician, Vol. 27 (1), pp. 17-21
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Machine Learning
A computer program is said to learn from experience E with respect to
some task T and some performance measure P, if its performance on
T, as measured by P, improves with experience E.
Ethics?ML & Ethics
A typical trade-off to be made in ML
is the minimisation of false positives (P1) or false negatives (P2).
“The decision must be based on values because there isn’t any purely
neutral and objective argument to support this choice.”
(Kraemer et al., 2011)
Kraemer, Felicitas; van Overveld, Kees; Peterson, Martin (2011), Is there an ethics of algorithms?, Ethics and Information Technology,
Sept. 2011, Vol. 13 (3).
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Machine Learning
A computer program is said to learn from experience E with respect to
some task T and some performance measure P, if its performance on
T, as measured by P, improves with experience E.
Ethics?ML & Ethics
Medical imaging: T = detecting infected cells
Avoid infected patient to be declared
not sick
“More false positive” consequences:
Side effects of useless surgery
Avoidable expenses for social security
P1: Minimizing false negative P2: Minimizing false positives
Avoid surgical intervention and
associated risks for healthy patients
Less avoidable expenses for social
security
“More false negative” consequences:
Infected patient not cured in time
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Well-known biaises of ML algorithms
• encode bias into data collection and analysis, hiding discrimination
behind seemingly objective numbers
• blind to effects they have in the world beyond the things
they are told to measure
• issues of resilience if decision algorithms are built on top of each
other and rely on the same interconnected data
• …
ML & Ethics
Source. Mulgan G. (2016), UK: how to grow informed public trust and maximise the positive impact of smart machines,
https://www.nesta.org.uk/sites/default/files/a_machine_intelligence_commission_for_the_uk_-_geoff_mulgan.pdf
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Outline
• Concepts
• Ethical issues in ML
• Ethical issues in Visualisation
• ML-supported visualisation
• Visu-supported ML
• Visu + ML
• Conclusion
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Visualisation & Ethics
Biases in visualisation
Bujack et al., TCVG 24(1), 2018
How to design a good continuous colormap?
Research is still needed…
We continue to map quantities to colours,
while colours can not be easily ordered,
are perceived in a (non-linear) way,
are subject to colour blindness issues…
Furthermore, cultural background also plays
some role in the interpretation of colours.
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Visualisation & Ethics
«To interpret data visualizations, people must determine how visual features
map onto concepts. For example, to interpret colormaps, people must
determine how dimensions of color (e.g.,lightness,hue) map onto quantities
of a given measure (e.g., brainactivity, correlation magnitude).»
(Schloss et al., 2019)
Karen B. Schloss et al., TCVG 25(1), 2019
Study on how inferred color-quantity mappings for colormap
data visualizations are influenced by the background color
Biases in visualisation
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Visualisation & Ethics
How to design a good scatter plot?
Research is still needed…
We continue to design scatter plots
with misleading scales,
with too many glyphs superimposed,
with multidimensional coding issues…
Sarikaya and Gleicher, TCVG 24(1), 2018
Recent work has shown that the order in
which scatter plots are shown has an
influence on class separability tasks
(Valdez, Ziefle and Sedlmair, TCVG, 2107)
Biases in visualisation
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Visualisation & Ethics
What if we take ethically sensitive decisions
based on a biased visualisations?
Biases in visualisation
Dimara et al., TCVG, 14(8), 2015
“The interplay between cognitive biases and visual data analysis
remains largely unexplored” Dimara et al., 2015
From a list of 154 cognitive biases reported in the literature,
Dimara et al., have built a list of 7 main categories:
• Estimation
• Decision
• Hypothesis Assessment
• Causal Attribution
• Recall
• Opinion Reporting
• Other
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Outline
• Concepts
• Ethical issues in ML
• Ethical issues in Visualisation
• ML-supported visualisation
• Visu-supported ML
• Visu + ML
• Conclusion
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can we use Machine-Learning
to design better Visualisations?
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can we use Machine-Learning to design better Visualisations? YES!
Requirements:
• IN {data, task(s), users}
• OUT { visualisations}
• Performance measure P (to be improved by
learning process)
Challenges:
• data, tasks, users can be described by several
variables of various types (no standards)
• visualisations can be described by several
variables of various types (many taxonomies)
• Performance measure P : quality of a
visualisation is very difficult to assess
(aethetics criteria, ease-of-use, effectiveness to
support a task, relationship to user skills and
domain background…)
Simplified ML Process
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can we use Machine-Learning to design better Visualisations? YES!
Simplified ML Process
ML to select best graph layout without drawing it
Kwon, Crnovrsanin, and Ma, TVCG 24(1), 2018
Fundamental assumption: «Given the same layout method, if the graphs have similar topological
structures, then they will have similar resulting layouts »
Training (supervised learning)
IN: graph data (topological structure) + layout
algo.
OUT: graph layout (aesthetic metrics)
Trained model: regression model between
(topological features & layout algo) and layout
results
New IN: new graph data + potential layout
algos.
OUT: estimation of aesthetic metrics
(without drawing the graph)
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can we use Machine-Learning to design better Visualisations? YES!
IN: Topological structure
• Graph kernel to measure pairwise similarities between graphs
• Graphlet frequencies as graph kernel (based on sampling)
• Graphlet frequency vector as the feature vector of a graph, then compute the similarity
between graphs by defining the inner product of the feature vectors.
Graphlets
Graphlets frequencies
Similar
graphlet
frequencies
Different graphlet frequencies
Kwon, Crnovrsanin, and Ma, TVCG 24(1), 2018
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can we use Machine-Learning to design better Visualisations? YES!
IN: Layout algorithms
• Force-directed method
• Dimension reduction based method
• Spectral method
• Multi-Level methods
• Clustering based methods
OUT: Aesthetic metrics
• Crosslessness: Minimizing the number of edge crossings
• Minimum angle metric: maximizing the minimum angle between incident edges on a
vertex
• Edge length variation: Uniform edge lengths
• Shape-based metric
Summary: ML approach using graph kernels to show different possible layouts of a given
graph and the related aesthetic metrics
Kwon, Crnovrsanin, and Ma, TVCG 24(1),
2018
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can we use Machine-Learning to design better Visualisations? YES!
Implicit Hypothesis in (classic) graph drawing:
one-to-one paradigm
an item (e.g. person) is represented by 1 node
a relation (e.g. friendship) is represented by 1 linkQuestion: why are some nodes duplicated?
Social interaction analysis
John Mary
Helen
one-to-one paradigm is not valid anymore
Biological pathway
An item can be displayed
many times in the pathway
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can we use Machine-Learning to design better Visualisations? YES!
Simplified ML Process
Training (supervised learning)
IN: graph data
for each node Ni: local topological features
around the node + global topological features +
domain related features of nodes*
OUT: binary var. telling if the node Ni has been
duplicated at later stages in the graph history
Trained model: SVM model
New IN: selected (new) node Nk to be displayed
in the biological pathway
OUT: suggestion to duplicate the node Nk
ML to suggest when nodes need to be duplicated in biological pathways visualization
ViBiNe project at LIST
Idea: learn from an history of manually curated and designed pathways
why some nodes have been duplicated
* e.g. type of biological entity
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Outline
• Concepts
• Ethical issues in ML
• Ethical issues in Visualisation
• ML-supported visualisation
• Visu-supported ML
• Visu + ML
• Conclusion
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can Visualisations help
to better use Machine-Learning?
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Simplified ML Process
Can Visualisation help to better use Machine-Learning? YES!
The ML process becomes the object to be visualized
(just like a business process or a biological mechanism)
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can Visualisation help to better use Machine-Learning? YES!
Consider ML as a process just like a business, industrial or biological process
• Data?
State of (internal) variables Visu. of Time-series?
Layer in Neural Network Visu. of (dynamic) Graphs?
Dimensions in reduced space… Visu. of Aggregated Data?
• Problem / Task?
Understand Model and Output
Diagnose False Predictions
Refine ML model
• Users?
Data scientists who want to fine tune ML algorithms
Business experts who need to justify the decision taken by algo.
…
Explanation, Interpretability of ML
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can Visualisation help to better use Machine-Learning? YES!
IEEE TCVG 25(1), 2019: Yao Ming, Huamin Qu, and Enrico Bertini RuleMatrix: Visualizing and Understanding Classifiers with Rules
Supporting domain expert to understand ML model
RuleMatrix system
Explain a black-box model by inducing a list of IF-THEN-ELSE rules
that can be visualised in interactive mode
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can Visualisation help to better use Machine-Learning? YES!
IEEE TCVG 25(1), 2019: Yao Ming, Huamin Qu, and Enrico Bertini RuleMatrix: Visualizing and Understanding Classifiers with Rules
Supporting domain expert to understand ML model
RuleMatrix system
Row Rule
Column Feature
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can Visualisation help to better use Machine-Learning? YES!
Supporting domain expert to understand ML model
Ribeiro et al., KDD 2016
Expert combines his/her (tacit) knowledge
and/or experience with the result of ML system
and takes decision on this basis
The model’s prediction is explained as a set of salient features
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can Visualisation help to better use Machine-Learning? YES!
Ribeiro et al., KDD 2016
Explaining predictions of competing classifiers
Correct for
wrong reason
Correct for
good reason
Supporting domain expert to understand ML model
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can Visualisation help to better use Machine-Learning? YES!
Explanatory Debugging, by Kulesza et al., IUI 2015
Auto. classification of messages in folders + Explanation of predictions
Folders
Messages of the selected folder Selected message
Why the message
has been
classified as such
List of words used by ML to make the prediction
Supporting domain expert to understand ML model
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can Visualisation help to better use Machine-Learning? YES!
Krause et al., VAST 2017
Supporting team of data scientist and expert to diagnose ML model
Model Diagnostics workflow is run after Model Building
ML Model seen as a blackbox, focus on input-output relationships
Better Understand ML models about how patients are handled in hospital:
Predict whether a patient coming to the emergency room will end up being admitted to the
hospital or sent home
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can Visualisation help to better use Machine-Learning? YES!
Predict if Patients will be admitted in the hospital based on the drugs they receive
Krause et al., VAST 2017Explanation Explorer
Item Level Inspector
Supporting team of data scientist and expert to diagnose ML model
16th April 2019Dr Ir Benoît Otjacques, [email protected]
IEEE TCVG 24(1), 2018: Wohngsuphasawat K et al. Visualising Data Graphs of Deep Learning Models in Tensor Flow
Can Visualisation help to better use Machine-Learning? YES!
Supporting data scientist to improve ML model
TensorFlow Graph Visualizer
Helping users to understand complex machine
learning architectures by visualizing their
underlying dataflow graphs
16th April 2019Dr Ir Benoît Otjacques, [email protected]
DeepEyes by Pezzotti et al., TCVG 2018 24(1)
Can Visualisation help to better use Machine-Learning? YES!
Visual Analytics to support the design of neural networks during trainingLoss and accuracy curves
Perplexity
histograms
Activation
Heatmap
Relationships
among the
filters in a layer
Filters activation
Supporting data scientist to improve ML model
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Can Visualisation help to better use Machine-Learning? YES!
VA to understand
hidden memories
of reccurent
neural networks
(in NLP)
RNNVis by Ming et al., IEEE VAST 2017.
Parameter
setting of RNN
Sentence visualisation
Word clusters
Hidden state clusters
Supporting data scientist to refine ML model
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Outline
• Concepts
• Ethical issues in ML
• Ethical issues in Visualisation
• ML-supported visualisation
• Visu-supported ML
• Visu + ML
• Conclusion
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Machine-Learning and Visualisationare complementary
Conclusions
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Conclusions
Why we need models that explain why they predict what they predict:
1. If AI < humans and not deployable identify the failure modes
(to guide further research)
2. If AI +/- = humans and deployable establish trust and
confidence in users
3. If AI > humans machine teach the humans about how to make
better decisions
Visualisation can play a role in the 3 cases.
Sozurce: Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, Dhruv Batra (2017),
Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization, ICCV 2017.
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Outline
Machine
LearningVisualisation
Use of ML to design Visualisations?
Use of Visualisations to better use ML?
Ethics? Ethics?
Ethics?
Ethics?
16th April 2019Dr Ir Benoît Otjacques, [email protected]
How to combine
a non-misleading visualisation
and a fair ML-based model
built from non-biased data
to support an ethically acceptable task?
Conclusions
16th April 2019Dr Ir Benoît Otjacques, [email protected]
Thank you for your attention
Benoît Otjacques
Feel free to contact me
if you believe that AI is a game changing technology,
that visualisation has a role to play to make AI more explainable,
and if you think that, whatever we do, we should keep ethics in mind.
I would also be happy to discuss with you if you disagree with these statements.