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


After some issues in the past…


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



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



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


Really?

Last week!


Outline

Machine

LearningVisualisation

Use of ML to design Visualisations?

Use of Visualisations to better use ML?

Ethics? Ethics?

Ethics?

Ethics?


Outline

• Concepts

• Ethical issues in ML

• Ethical issues in Visualisation

• ML-supported visualisation

• Visu-supported ML

• Visu + ML

• Conclusion


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.


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


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)


Concepts

Information

Visualisation

Visual

Perception

Computer

Vision

Computer

GraphicsInfographics

Visual

Analytics

Scientific

Visualisation

Informative

Art


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


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


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


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


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


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


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


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


Concepts

There are still other concepts…

Data Visualization

Visual Language

Graphic Design

Visual Communication

Visual Thinking

Knowledge Visualisation

…


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?


• Concepts





• Visu + ML

• Conclusion

Outline


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?


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


When you see the pictures, is it so difficult to realize that the training dataset is biased?

ML & Ethics


Machine Learning




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?


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


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


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


Machine Learning




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."


Machine Learning




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


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


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).


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


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


Outline

• Concepts





• Visu + ML

• Conclusion


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.



«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




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)




What if we take ethically sensitive decisions

based on a biased visualisations?


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


Outline

• Concepts





• Visu + ML

• Conclusion


Can we use Machine-Learning

to design better Visualisations?


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




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)



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



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



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




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


Outline

• Concepts





• Visu + ML

• Conclusion


Can Visualisations help

to better use Machine-Learning?



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)



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



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



IEEE TCVG 25(1), 2019: Yao Ming, Huamin Qu, and Enrico Bertini RuleMatrix: Visualizing and Understanding Classifiers with Rules


RuleMatrix system

Row Rule

Column Feature




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



Ribeiro et al., KDD 2016

Explaining predictions of competing classifiers

Correct for

wrong reason

Correct for

good reason




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




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



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


IEEE TCVG 24(1), 2018: Wohngsuphasawat K et al. Visualising Data Graphs of Deep Learning Models in Tensor Flow


Supporting data scientist to improve ML model

TensorFlow Graph Visualizer

Helping users to understand complex machine

learning architectures by visualizing their

underlying dataflow graphs


DeepEyes by Pezzotti et al., TCVG 2018 24(1)


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



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


Outline

• Concepts





• Visu + ML

• Conclusion


Machine-Learning and Visualisationare complementary

Conclusions


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.


Outline

Machine

LearningVisualisation

Use of ML to design Visualisations?

Use of Visualisations to better use ML?

Ethics? Ethics?

Ethics?

Ethics?


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


Thank you for your attention

Benoît Otjacques

[email protected]

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.

Visualisation as a mean to tackle some ethical issues ... · Visualisation as a mean to tackle some...

Documents

Transcript of Visualisation as a mean to tackle some ethical issues ... · Visualisation as a mean to tackle some...