Dr. Marcel Blattner - The Black Box Problem
The Black Box Problem 1
Dr. Marcel Blattner
Chief Data Scientist TX Group
marcel.blattner@tx.group

@blattnerma
Dr. Marcel Blattner - The Black Box Problem
The Black Box Problem 2
5 Ways Artificial Intelligence Will Change the World by 2050
https://news.usc.edu/trojan-family/five-ways-ai-will-change-the-world-by-2050/
How artificial intelligence will change every aspect of our lives
https://nypost.com/2018/09/07/how-artificial-intelligence-will-change-every-aspect-of-our-lives/
How Artificial Intelligence Is Totally Changing Everything
https://science.howstuffworks.com/artificial-intelligence.htm
AI will write a best-seller by 2049, experts predict

https://www.weforum.org/agenda/2018/03/timeline-of-creative-ai/
- What have opinion leaders to offer?
X → YWe can summarise these statements: Such statements are
useful only if X is
understood! Sadly,
X is not enough
understood in most
cases
Hypeitup!OMG!!
Dr. Marcel Blattner - The Black Box Problem
The Black Box Problem 3
- Who are the opinion leaders?
• Nearly 60 percent of news articles across industries are
indexed to industry products, initiatives, or announcements.
33 percent of unique sources across all articles are affiliated
with industry, almost twice as many as those from
academia, and six times as many as those from
government. Nearly 12 percent of all articles reference Elon
Musk.
You can safely count me in here!
https://reutersinstitute.politics.ox.ac.uk/our-research/industry-led-debate-how-uk-media-cover-artificial-intelligence
That’s bad news. Opinion
in public is driven

mainly by completely
biased
industry players
Dr. Marcel Blattner - The Black Box Problem
The Black Box Problem 4
“…..the dynamics of learning observed in deep neural networks
remain much of a mystery to this day.”
- What does a giant (Yoshua Bengio) say about AI and ANN?
On the Learning Dynamics of Deep Neural Networks (2019)
I have some questions!
Dr. Marcel Blattner - The Black Box Problem
5
Input Output
You don’t know the inner working.
Usually we do not care about using black boxes!
Why?
The Black Box Problem
- What we mean by Black Box
Dr. Marcel Blattner - The Black Box Problem
What is a black box? 6
- How to make a computer smart?
Data New
Data
Deductive
Algorithms
Model
Abstraction
Inductive
Training Application
Connectionists
• Rules are based on abstractions created by machines

• Rules are mostly not interpretable by humans 

• Expressive advantages 

• System stability and general behaviour is not

fully predictable
Algorithms
Experten-

Knowledge
Application
Rules
New
Data
Deductive
if/then

else
Symbolic
• Rules are human knowledge based

• Rules are transparent

• Expressive disadvantage
Dr. Marcel Blattner - The Black Box Problem
7
The Black Box Problem
- Object-Detector (CNN)
- Trained on 1.8 Mio. images
- Ported to iOS for real time

object recognition
- How to confuse a Artificial Neural Network
Dr. Marcel Blattner - The Black Box Problem
8
Source: Adversarial Patch: https://arxiv.org/pdf/1712.09665.pdf
Adversarial Attack - example
The Black Box Problem
- How to confuse a Artificial Neural Network
Dr. Marcel Blattner - The Black Box Problem
9
The Black Box Problem
Making a Tesla 50mph faster than allowed
https://electrek.co/2020/02/19/tesla-autopilot-tricked-accelerate-speed-limit-sign/
- How to confuse a Artificial Neural Network
Dr. Marcel Blattner - The Black Box Problem
10
The Black Box Problem
- What is needed to make a black box interpretable?
InterpretabilityTransparency
Based on
meta theories, e.g.

real understanding of the learning
dynamics
Explainability
Based on post-hoc

experiments, e.g.

increase experience of 

input-output relation
Operator Executor
Decision
subject
Creator
According Tomsett et al. (2018)
Dr. Marcel Blattner - The Black Box Problem
11
The Black Box Problem
- Explainability
Explainability
Based on post-hoc

experiments, e.g.

increase experience of 

input-output relation


Layer-Wise Relevance
Propagation
pug, pug-dog (254): 18.939
French bulldog (245): 11.256
bull mastiff (243): 10.136
doormat, welc ome mat (539): 9.809
Norwegian elk hound, (174): 9.591
Dr. Marcel Blattner - The Black Box Problem
12
Theoretical framework
ANN
(specific architecture)
Mapping to a framework
(possibly outside CS)
Predictions /
new insights
Guided experiments
Inverse mapping
Hypothesis
Proposal
Testing by experiments
with ANN
The Black Box Problem
- Transparency
Dr. Marcel Blattner - The Black Box Problem
Theoretical framework
ANN
(specific architecture)
Mapping to a framework
(possibly outside CS)
Metastable
configurations
(mini/max Landscape
Guided experiments
Inverse mapping Hypothesis
Loss Surface
Testing by experiments
with ANN
H = −
∑
i1>⋯>ip=1
Ji1...ip
σi1
⋯σip
Y = q
n0
∑
i=1
γ
∑
j=1
Xi, j Ai, j
K
∏
k=1
w(k)
i, j
13
- What could be a good candidate for a theory?
The Black Box Problem
TheorySimulations
Dr. Marcel Blattner - The Black Box Problem
Summary
14
• The Black Box Problem is important because of the

technology adoption in industry, government and private
sectors
• Try harder!! We need better models (theories) to
understand these technologies
• Theoretical physics could make valuable 

contributions

Ai black box

  • 1.
    Dr. Marcel Blattner- The Black Box Problem The Black Box Problem 1 Dr. Marcel Blattner Chief Data Scientist TX Group marcel.blattner@tx.group
 @blattnerma
  • 2.
    Dr. Marcel Blattner- The Black Box Problem The Black Box Problem 2 5 Ways Artificial Intelligence Will Change the World by 2050 https://news.usc.edu/trojan-family/five-ways-ai-will-change-the-world-by-2050/ How artificial intelligence will change every aspect of our lives https://nypost.com/2018/09/07/how-artificial-intelligence-will-change-every-aspect-of-our-lives/ How Artificial Intelligence Is Totally Changing Everything https://science.howstuffworks.com/artificial-intelligence.htm AI will write a best-seller by 2049, experts predict
 https://www.weforum.org/agenda/2018/03/timeline-of-creative-ai/ - What have opinion leaders to offer? X → YWe can summarise these statements: Such statements are useful only if X is understood! Sadly, X is not enough understood in most cases Hypeitup!OMG!!
  • 3.
    Dr. Marcel Blattner- The Black Box Problem The Black Box Problem 3 - Who are the opinion leaders? • Nearly 60 percent of news articles across industries are indexed to industry products, initiatives, or announcements. 33 percent of unique sources across all articles are affiliated with industry, almost twice as many as those from academia, and six times as many as those from government. Nearly 12 percent of all articles reference Elon Musk. You can safely count me in here! https://reutersinstitute.politics.ox.ac.uk/our-research/industry-led-debate-how-uk-media-cover-artificial-intelligence That’s bad news. Opinion in public is driven
 mainly by completely biased industry players
  • 4.
    Dr. Marcel Blattner- The Black Box Problem The Black Box Problem 4 “…..the dynamics of learning observed in deep neural networks remain much of a mystery to this day.” - What does a giant (Yoshua Bengio) say about AI and ANN? On the Learning Dynamics of Deep Neural Networks (2019) I have some questions!
  • 5.
    Dr. Marcel Blattner- The Black Box Problem 5 Input Output You don’t know the inner working. Usually we do not care about using black boxes! Why? The Black Box Problem - What we mean by Black Box
  • 6.
    Dr. Marcel Blattner- The Black Box Problem What is a black box? 6 - How to make a computer smart? Data New Data Deductive Algorithms Model Abstraction Inductive Training Application Connectionists • Rules are based on abstractions created by machines • Rules are mostly not interpretable by humans • Expressive advantages • System stability and general behaviour is not
 fully predictable Algorithms Experten-
 Knowledge Application Rules New Data Deductive if/then else Symbolic • Rules are human knowledge based • Rules are transparent • Expressive disadvantage
  • 7.
    Dr. Marcel Blattner- The Black Box Problem 7 The Black Box Problem - Object-Detector (CNN) - Trained on 1.8 Mio. images - Ported to iOS for real time
 object recognition - How to confuse a Artificial Neural Network
  • 8.
    Dr. Marcel Blattner- The Black Box Problem 8 Source: Adversarial Patch: https://arxiv.org/pdf/1712.09665.pdf Adversarial Attack - example The Black Box Problem - How to confuse a Artificial Neural Network
  • 9.
    Dr. Marcel Blattner- The Black Box Problem 9 The Black Box Problem Making a Tesla 50mph faster than allowed https://electrek.co/2020/02/19/tesla-autopilot-tricked-accelerate-speed-limit-sign/ - How to confuse a Artificial Neural Network
  • 10.
    Dr. Marcel Blattner- The Black Box Problem 10 The Black Box Problem - What is needed to make a black box interpretable? InterpretabilityTransparency Based on meta theories, e.g.
 real understanding of the learning dynamics Explainability Based on post-hoc
 experiments, e.g.
 increase experience of 
 input-output relation Operator Executor Decision subject Creator According Tomsett et al. (2018)
  • 11.
    Dr. Marcel Blattner- The Black Box Problem 11 The Black Box Problem - Explainability Explainability Based on post-hoc
 experiments, e.g.
 increase experience of 
 input-output relation 
 Layer-Wise Relevance Propagation pug, pug-dog (254): 18.939 French bulldog (245): 11.256 bull mastiff (243): 10.136 doormat, welc ome mat (539): 9.809 Norwegian elk hound, (174): 9.591
  • 12.
    Dr. Marcel Blattner- The Black Box Problem 12 Theoretical framework ANN (specific architecture) Mapping to a framework (possibly outside CS) Predictions / new insights Guided experiments Inverse mapping Hypothesis Proposal Testing by experiments with ANN The Black Box Problem - Transparency
  • 13.
    Dr. Marcel Blattner- The Black Box Problem Theoretical framework ANN (specific architecture) Mapping to a framework (possibly outside CS) Metastable configurations (mini/max Landscape Guided experiments Inverse mapping Hypothesis Loss Surface Testing by experiments with ANN H = − ∑ i1>⋯>ip=1 Ji1...ip σi1 ⋯σip Y = q n0 ∑ i=1 γ ∑ j=1 Xi, j Ai, j K ∏ k=1 w(k) i, j 13 - What could be a good candidate for a theory? The Black Box Problem TheorySimulations
  • 14.
    Dr. Marcel Blattner- The Black Box Problem Summary 14 • The Black Box Problem is important because of the
 technology adoption in industry, government and private sectors • Try harder!! We need better models (theories) to understand these technologies • Theoretical physics could make valuable 
 contributions