Full day workshop delivered at IA Conference 2019.

1. AI for Information Architects
User Experience, Content Strategy and
Design
1
Marianne Sweeny
IA Conference 2019 Workshop

2. 2
During our day
together we will
examine
What constitutes intelligence
Consciousness as part of the human
experience
Differences between Machine Learning and
Artificial Intelligence
Where did AI come from?
AI issues: safety, ethics and privacy
The intersections of AI, IA, user experience
and content strategy
How to use this knowledge to design for
humans and machines

3. Link to Workshop PDF
Tiny URl

4. Instructions for Living a Life
Pay attention
Be astonished
Tell about it
4
Mary Oliver

5. Foresight and Hindsight
All technological change is a trade-off
The advantages and disadvantages of new
technologies are never distributed evenly among the
population
Embedded in every technology there is a powerful
idea, sometime two or three ideas
Technological change is not additive; it is ecological
Media tend to become mythic (Computationalism)
5

6. Why Is This Important?
Because they are developing an GUI for AI
that you will be able to use to build artificially
intelligent interactions 6

7. Why is This Important? (2)
Because, unlike the Manhattan Project,
there is no governance over who is
doing what.
7

8. Why is this Important? (3)
Because AI is not infallible yet the
consequences are forever
8

9. Why is This Important (4)
Because it is our job as information architects,
user experience professionals, content
strategists, human factors professional
9

10. Why is This Important (5)
Quantum Around the Corner
Google said it had already devised machine-learning
algorithms that work inside the quantum computer,
which is made by D-Wave Systems of Burnaby,
British Columbia. …The most effective methods for
using quantum computation, Google said, involved
combining the advanced machines with its clouds
of traditional computers.

11. INTELLIGENCE
11

12. Mai Quality of Information
Information (Intelligence) is part of a spectrum
Data >> Information >> Knowledge
Information quality depends on individual
characteristics
– Contextual
– Situational
– Environmental
– Emotional
Machines use captured personal data

13. Reasoning
Deductive
Theory
Hypothesis
Observation
Confirmation
Inductive
Theory
Hypothesis
Pattern
Observation

14. Biological System Levels of Reasoning
Computable outcome (goal)
Steps/instructions to realize outcome (algorithm)
Implementation of program (realization of goal)
14

15. Artificial Meaning
“Context has always been part of expression because
expression become meaningless if context becomes
arbitrary…meaning is only ever meaning(ful) in
context.
…
Any gadget, even a big one like Singularity, gets
boring after awhile. But a deepening of meaning is
the most intense potential kind of adventure available
to us.”
15

16. Artificial Context
Context becomes what the system can measure
• Environmental features
• Interactions
• Ubiquitous computing
• Internet of things (IoT)
Non-methodical approach that brings in containment
(social through local) interactions
• Adaptive/reactive interaction in situ
• Context as perceived and used by actor
16

17. Information Cascade
A group of agents behaving rationally can fall prey to
infinite misinformation
• US Vaccination controversy
Information Cascade: when rational theory is based
on filter bubbles, hive mind
Cascade is caused by a misinterpretation of what
others think based on external observation of their
actions
More concerned with judgement fitting existing
consensus than the visible facts
17

18. Intelligence Explosion
Human-level AI will lead to super human AI
• Uncontrolled intelligence explosion without
human-level intentionality that is the result of
consciousness
• Program self-improves to state that exceeds ability
for outside control
Intelligence here measured by ability to attain
goal in most efficient manner
18

19. Information Explosion Components
Components
• Increased computational resources
• Duplicability
• Editability
• Goal Coordination
Accelerators
• Hardware capacity
• Better algorithms
• Massive datasets
• Psychology and neuroscience applications
• Accelerated science (quantum computing)
• Economic incentives (labor $ reduction)
19

20. Computationalism
World can be understood by computational processes
with humans as sub processes
1st Flavor: logical positivism
2nd Flavor: computer program with features related
to self representation and circular references similar to
that of a person
3rd Flavor: information structure that can be
perceived by some real human to also be a person
(Turing Test)
20

21. Solutionism
Silicon Valley assumption of a quantifiable self that is the
truer self
There’s an app for everything
False notion that Internet is a coherent and stable
influence in our lives
Grasping easy digital solutions often ignores complex
causes behind
Sometimes right algorithms can lead to wrong answers
21

22. Subjectivism
John Searle: AI not possible in any way because
consciousness is a physical property of the brain that
produces a subjective experience
Thomas Nagal: computers do not have subjectivity
(private landscape with personal experience). Cannot
create
Subjective Reality
• Intangible way to intelligence
• Philosophical concept focused on sense of self and
components (experience, perspective, belief, emotion,
consciousness)
• Composed of understanding and intentionality
Introspection is key
22

23. 23
Consciousness

24. Gelernter: Tides of Mind
Humans have a knowledge of core concepts related to
the physical world = consciousness
Consciousness allows for building more robust mental
models that enable inference and prediction
Key question going unanswered: What is the human
mind without the human being?
The mind is consciousness (objects & events) plus
memory (occurred outside of the mind)
Thinking has intuitive meaning tied to consciousness
• Perception
• Recollection
• Idea
24

25. Gelernter on Consciousness
“Conscious experiences range from vivid color
sensations to experience of the faintest
background aromas; from hard-edged pains to
the elusive experience of thoughts on the tip of
one’s tongue. . . . All these have a distinct
experienced quality. . . . To put it another way,
we can say that a mental state is conscious if it
has a qualitative feel—an associated quality of
experience…”
25

26. Consciousness Spectrums
Up-spectrum
• Live in the present
• Outer consciousness: external world (bodies)
• Feeds memory up
• Thinking is focused, disciplined systematic
Down-spectrum
• Recall, revisit, reoccupy the past
• Dreams are re-experiencing memories in the form of
thought
• Recollections, ideas
26

27. Magical Thinking
Magical thinking = things only imagined become real
Cannot “learn” to be creative
Creativity is repurposing in a way that software cannot
because it involves:
• Ignoring limits
• Curiosity
Inspire but not force creative insight
“Where the confines of the waking world blend with
those of dreams.” Edgar Allen Poe
27

28. Dreams
“Dreams tell us truths that we know but are not brave
enough to acknowledge.”
Remembering out of control
Dreams = emotions + hallucinations
28

29. 29
For too long, emotionhas been cognitive researchers’
third rail. In research on humans, emotions were
deemed irrelevant,impossible to study or beneath
scientific notice…But nothing could be more essential to
understanding how people and animals behave.
Sy Montgomery, NYT Mar 3 2019

30. Emotion
Primary emotions
• Interest
• Pleasure
• Distress
Secondary emotions
• Anger
• Fear
• Disgust
• Happiness
• Sadness
• Surprise
Emotional Resonance: ability to feel/echo someone
else’s feeling
• Empathy
• Sympathy
• Essential to the human experience
30

31. Emotions (2)
Play key role in decision-making, creativity and
intelligence (EQ)
Sentic Modulation
• Facial Expression
• Voice: utterance, timing, pitch
Learning is the quintessential learning experience
31

32. Affective (emotional) Computing
Assumption of small set of emotions to make
programming easier
Assumes binary nature of emotions (cannot be angry and
pleased)
Conversational signals
• Syntactic displays
• Speaker Displays
• Listener Response displays
Emotionally-basedcomputers
• Same emotional ability similar to a dog, neither personal or friendly
• Computer voice with intonation and natural expression
• Computer perceives emotional state and responds appropriately
• Maximized sentic communication between human and computer,
personal and “user-friendly”
32

33. MACHINE LEARNING
Capacity for system to improve
performance through experience, not
explicit instructions
33

34. Machine Learning: Programs that act like humans
34

35. Machine Learning
A programming approach to problem-solving– composite
of not a single algorithm
Model of real world using mathematic structure with
decision-makingrules
Derives rules from a data set
Objective function = desiredoutcome
Training set with adjusted parameters until goal achieved
Test set used to validate accuracy and effectiveness
Machine completes an objective without specific instructions
35

36. 36

37. Unsupervised Learning
Unlabeled data
Clustering
Segmentation
Association
Algorithms
• Neural networks
• Independent component analysis
37

38. Supervised Learning
Uses document-class pairs to indicate proper classes for
given documents
Used human specialists for classification of “training set”
used to “teach” system
• Assigns classes to documents
• Reviews machine classification performance
6 Algorithm types
• Decision Trees
• Nearest neighbor
• Relevance Feedback
• Naïve Bayes
• Support Vector Machine
• Ensemble
38

39. Probabilistic Machine Learning
Probabilistic framework can represent and manipulate
uncertainty
Requires high capacity for flexibility to allow data to
“speak for itself”
Universal inference engine using Monte Carlo
39

40. Reinforcement Learning
Program learns reward from human feedback then
optimizes reward function
• Rewards
o Sampled
o Evaluative
o Sequential
• Optimized reward function
• Reward must be explicit to avoid being “gamed”
Issues with tasks and goals
• Too complex
• Hard to specific
• Poorly defined
40

41. Transfer Learning (CS)
Reason relationally
Requires conceptual representation produced by
abstract structural knowledge (that is where we humans
come in)
Generalizations are transferred to environments that
share structures, e.g. mental models
41

42. ARTIFICIAL INTELLIGENCE
Solve problems that the mind can solve using
derived intelligence instead of a prearranged set of
rules
42

43. Generalized Intelligence
Spearman coefficient to measure intelligence,
correlation measure, if/then
G Factor: general level of intelligence possessed by an
individual
Quantified intelligence represented by a number
Used to rank people by IQ
43

44. 44

45. Neuroscience of Algorithms
Deep Learning
• Distributed interactions
• Tuned by learning procedures
• Stochastic (random) parallel
information processing
Convoluted neural
networks
• Convergent and divergent
information flows
• Non linear transduction
• Maximum-based pooling of
inputs
45

46. Semantic Computing
Segmented and match instruction
Associations to understand human behavior and
predict actions
Requires semantic matching
• Control layer (input)
• Semantic mapping layer (ontology)
• Device Layer
Requires user and behavior models (persona)
Semantic reasoning module confirms user intentions
46

47. Deep Learning
47

48. Deep Learning Components
Collection of trainable math units which collaborate to
compute complicated functions
HUGE raw data training set
Results get better with more data, new/better algorithms
based on observation and insight
Requirements
• Scalable
• Portable
• Reproducible
• Extensible
• Powerful processing hardware
48

49. AI Types
Logical Reasoning
Knowledge Representation
Planning and Navigation
Natural Language Processing
Perception
49

50. Embodied Agents
Internet of Things
Goal driven planning
Reactive agents
Search
50

51. What AI Best Suited To
Search
Learning Systems
Pattern Recognition
Planning
Induction
51

52. Search
Requires additional structure
Near to | Close to expansion
Solve for one, Solve for many
Personalization
52

53. Learning Systems
Use past behavior to predict future action using
human planned heuristic methods
A reinforced learning model that leads to a secondary
reinforcement model that is more autonomous
• Reinforcement is reward
• Extinction is unlearning
Grade on curve of computer’s acquired capability
53

54. Pattern Recognition
Ability for computer to act intelligently based on input
data with a lot of variability
• Decision Trees
• Nearest neighbor classification
• Neural Networks
Classification
Ideal replaced by practical
54

55. Planning & Problem-Solving
Large assembly of interrelated sub-problems
Given a start state and desired outcome state
Choose appropriate sub-problems for solving selected
problem
Success is most efficient set of actions to achieve
desired outcome
55

56. Induction
Learning by example
Derive the rule from set of observed instances
Classification key component
• A learning system has to be capable of evolving its own
class descriptions
• The task of constructing class definitions is called
induction or concept learning
56

57. AI Models
Base Models
• Learning
• Prediction: create actions to respond to learning
Sub modules
• Data analysis
• User Identification
• Behavior recognition
• Service construction and provisioning
57

58. Thought Vectors
Gregory Hinton – Google Research Fellow
Encode thoughts as sequences of numbers (vector)
Software learns to recognize patterns in these digital
representations
“If you take the vector for Paris and subtract
the vector for France and add Italy, you get
Rome,” he said. “It’s quite remarkable.”
Geoffrey Hinton
58

59. Google BERT
AI to carry on decent conversation (Turing test)
Learn general vagaries of language and apply to
specific task
Analyzed millions of sentences
• Self-published literature
• Entire Wikipedia
Goal to predict next word and understand the
fundamental relationships between words
59
Bio directional Encoder Representations and Tranformer Language

60. Singularity
State when humans and machine merge
Concept introduced by Ray Kurzweil, Google fellow
6 Epochs of Evolution
• Physics & Chemistry
• Biology
• Brains
• Technology
• Merger of Technology & Human
Intelligence
• Universe Integration
60

61. Super Intelligence
Traits
• Capacity to learn
• Capacity to deal with uncertainty
• Ability to extract concepts from data and internal state
• Ability to leverage acquired concepts for combinatorial
representations for logical & Intuitive reasoning
• Capacity for unrestrained self-improvement (overwrite its own
code)
Types
• Speed (faster than human mind)
• Quality (faster and much smarter than human)
• Collective (aggregates performance of lesser intelligences)
External governance: None
61

62. And We Go Boldly Into the Whirling Knives*
AI might achieve a strategic advantage
Orthogonality Thesis: cannot assume that AI would be
able to share our biological values
• Culture
• Kindness
• Spiritual enlightenment
Instrumental Convergence Thesis: cannot assume that
Super AI would be satisfied with a supportive or
subservient role
Super AI could develop a final goal that is not
anthropomorphic
62
Existential Risks

63. Whirling Knives (2)
Perverse Instantiation: satisfy goal in a way that
violates programmed intent
Infrastructure profusion: over consumes resources to
achieve more reward
Mind crime: AI creates processes with moral states
(sentient simulations)
63
Practical Risks

64. 64

65. Break
https://affinelayer.com/pixsrv/

66. ORIGINS OF AI
66

67. Turing (1948)
Imitation Game
Reward Signal increases probability of repetition events
leading to it
Suitable imperative: one that regulates the order in
which the rules of the logical system are applied
Program the learning
Objections to machine intelligence
• Theological
• Head in Sand
• Mathematical/philosophical
• Consciousness
• Lady Lovelace
67

68. Dartmouth Summer Research Project (1956)
Top AI scientists proposed
concentrated effort for AI
Focused on
• Computer use of language
• Neuron Nets
• Self improvement
• Abstraction
• Randomness and creativity
Follow up conference held in
2006
68
Human Factors professionals NOT INVITED

69. Perceptron (1957)
Developed by Frank Rosenblatt, psychologist at Cornell
Artificial neural networks (ANN) = many interconnected
processing units for parallel processing
Trained not programmed
1. Input addition
2. Comparison with threshold value
3. If threshold met or surpassed, output activation
Modifiable connections adjusted according to “learning”
algorithm
Perceptrons are not without limitations
69

70. Pandemonium (1959)
Decision making entity involving 4 “demons”
Top Layer decides what information has been presented
to the system (discernment)
Feed-forward and feedback connections between layers
Requirements:
• Well defined problem
• Unbiased decision making
• Single tamper-proof labeling of behavior
70

71. Pandemonium Feed Forward Layers
Bottom layer: store data
3rd Layer: select, weigh, filter and pass along data
2nd Layer: “cognitive demons” decide which
information from 3rd layer to process
Decision layer: single decision demon on what
information is presented to the system for
processing
71

72. Marvin Minsky (1960)
Cognitive computer scientist
Co-founder MIT AI Laboratory
Symbolic AI
With Seymour Papert brought forth a 20 year “AI
winter” with criticism of early AI Artificial Neural
Network (ANN) approach
72

73. Minsky on What AI Best Suited To
Search
Learning Systems
Pattern Recognition
Planning
Induction
73

74. Minsky on Creativity
“There’s no such thing as “creativity” in the first place. I
don’t believe there’s any substantial difference
between ordinary thought and creative thought…I’ll
argue that this is really not a matter of what’s in the
mind of the artist – but what’s in the mind of the
critic…”
74

75. ELIZA (1963)
1st Instance of human mediated chatbot
Early computer/human conversation (NLP)
Heuristic programming
• Keyword identified (input)
• Sentence transformed according to rule associated with
identified keyword
• Choose appropriate transformation – if none available,
choose most likely/earlier transformation
• Generate responses
Keyword dictionary contains composition, assembly
and decomposition rules
75

76. Eliza and Global Context
Global Context is key to understanding
Sub-contexts emerge as conversation continues for
consequential richness
Individual participants bring their own belief
structure
ELIZA scripts (previous learnings) establish a global
context for future “understandings”
Broad context framework only
76

77. Dreyfus (1964)
Create computer systems with the intelligence and
reasoning of an human adult
Rationalist assumption of “ordered reality” is flawed
Knowledgeable reality itself lacks rational structure
Inter-relatedness between humans and the world
Human world filled with experience structures- neither
subjective or objective
AI discovers meaningful structures to apply to
meaningful behaviors, independent of fixed rules
77

78. Norvig & Russell (2004)
Systems that act like humans (Turing)
System that think rationally (logic solvers)
Systems that act rationally (perception, NLP,
Planning, Navigation)
Systems that think like humans (neural)
78

79. Sloman Minsky Model (2000)
79

80. St Thomas Symposium (2004)
Need synthesis of methodologies
Move from reactive to deliberative thinking
Include affective concepts like emotions
• Primary
• Secondary
• Tertiary
Incorporate “common sense” thinking
Source of human resourcefulness and robustness
80
Human Factors professionals NOT INVITED

81. Asilomar Conference 2009
81
Human Factors professionals NOT INVITED

82. Asilomar AI Research Needs: HCI
82
Human Factors professionals NOT INVITED

83. Stanford 100 Year AI Study (2016)
Long term reoccurring study of AI influence on people
and society
Modeled after Association for Advancement of Artificial
Intelligence (AAIA) consortium 2008
4 intended audiences
• General public
• Industry
• Government
• AI Researchers
83
Human Factors professionals NOT INVITED

84. 5 Dominant Tribes of AI (2017)
Symbolists: logical reasoning
Connectionists: structures inspired by human brain
Evolutionaries: methods inspired by Darwin theory of
evolution
Bayesians: probabilistic inferences (google and others)
Analogizers: extrapolate from previously seen
examples
84

85. Two Schools of AI
Symbol Processing
Neural nets
85

86. Symbolic AI
Intelligence = symbol manipulation
Fixed and formal rules
Assume: all intelligent processes are forms of
information processing
Computer processes symbolic representations (1s/0s)
according to formal rules (program)
Plato’s rationalism
GOFAI
86

87. Artificial Neural Networks
Connectionism
Neural networks made up of input layer, interstitial
layers and output layer
Good at: pattern recognition, categorization, and
behavior coordination
Knowledge comes from the connections not symbol
interpretation
Past experience used to form intelligence in current
state
Heideggerian AI
87

88. 88

89. Learning Systems
Use past behavior to predict future action using
human planned heuristic methods
Reinforced learning model that leads to a secondary
reinforcement model that is more autonomous
• Reinforcement is reward
• Extinction is unlearning
Grade on curve of computer’s acquired capability
89

90. Pattern Recognition
Ability for computer to act intelligently based on input
data with a lot of variability
• Decision Trees
• Nearest neighbor classification
• Neural Networks
Classification
Ideal replaced by practical
Constant decision what problem to work on
• Value based
Pandemonium
90

91. Planning & Problem-Solving
Large assembly of interrelated sub-problems
Choose appropriate sub-problems for solving selected
problem
Logic Theory: prove theorem using heuristics:
• Similarity test
• Simplicity test
• Strong non-provability test
Heuristic programming
91

92. Heuristic Programming
Early training for AI
Self-learning
• Substitutes machine learning for logic algorithms
• Ranks alternative in a branching decision trea
Achieves an approximate of the exact solution
ELIZA
92

93. Optimal Stopping
Computer science problem
Stop too early and you miss a good candidate
Stop too late and you miss a good candidate waiting
for perfection that doesn’t exist
Threshold rule: establish a optimal stopping point and
take the first candidate above that percentile
Establish a “period of no decision” – predetermined
amount of time for looking then a leap phase of
commit
93

94. Explore | Exploit Tradeoff
Explore: gathering information
Exploit using the information gathered to produce a
good result
Value of explore declines over time
Value of exploit increases over time
Exploration has inherent value of finding the best
candidate
“To live in a restless world requires a certain restless in
oneself…you must never fully cease exploring.” p.54
94

95. The Principle of Beneficence
Philosophic concept tied to ethics
Condition of “do no harm” in medicine
Possible harm to few in order to benefit many
Philippa Foot and moral dilemmas (train switch
scenario)
Who decides the winners and losers of AI?

96. LUNCH BREAK 96

97. AI ISSUES
97
The most exciting phrase to hear in
science, the one that heralds new
discoveries, is not ‘Eureka’ but “that’s
funny…” Isaac Asimov

98. Learning is one Thing…Thinking
Another
“In designing software and microprocessors, I
have never had the feeling that I was designing
an intelligent machine. The software and
hardware is so fragile and the capabilities of the
machine to “think” so clearly absent that even
as a possibility, this has always seemed very far
in the future…My personal experience suggest
we tend to over estimate our design abilities.”
98

99. Sometimes They Learn the Wrong
Things
99

100. Sometimes They Get Things Wrong
100

101. Sometimes They Do the Wrong Thing
101

102. Sometimes They Build the Wrong Things?
Built as a proof of concept
for AI gone wrong with
biased data
MIT AI Lab
Dataset was a sub-reddit
dedicated to document the
“disturbing reality of death.”
102

103. 103

104. 104
Privacy

105. User Metrics Training Data
Frequency of access
Click-through (selection from results set)
Time on site
Pages per session
Bounce Rate
Conversion (fulfilled information need)
Profile data
105

106. Implicit Collection
Implicit (max precision 58%)
• Software agents
• Logins
• Enhanced proxy servers
• Cookies
• Session IDs
Gathered without user awareness from behavior
• Query context inferred
• Profile inferred
• Less accurate
• Requires a lot of data
106

107. Explicit Collection
Explicit (max precision 63%)
• HTML forms
• Explicit user feedback interaction (early Google
personalization with More Like This)
Provided by user with knowledge
More accurate as user shares more about query intent
and interests
107

108. What Constitutes a User Profile
Information types
• Demographic
• Interests (short & long-term)
• Preferences
Profiles are dynamic and iterate over time
Represented as
• Set of weighted keyword
• Weighted concepts
• Semantic network
108

109. Google on Privacy (2007)
“There was a small trade off on privacy but they’re
going to get dramatically better search results. That
was something that made sense to us over time.”
Melissa Mayer
VP User Experience
Google
109

110. Google on Privacy Now (2019)
https://twitter.com/jason_kint/status/11054840
10183188480 110

111. What Google Collects
Implicit
Use information
Device information
Log information
Unique application information
Local storage
Cookie data
Explicit
Location information
Profile information
111

112. What Google Collects
Profile information
Use information
Device information
Log information
Location information
Unique application information
Local storage
Cookie data
112

113. Methods
Client Side: gather data from user profile
Server Side: gather data from system usage (logs)
Group-ization: Recommender system with vested
interest
Member data used to rank the individual results
• Relevance weigh enhanced with more members of
group who “like” resource
• Sum of personalization scores of each group member
113

114. Google Personalization
Tracks
• What is selected
• Level of interaction
• What is not-done
(bounce rate)
Signals
• Location
• Search history
Less specific queries
benefit the most as they
require the additional
context provided by
personalization
114

115. Facebook Security Lapses
2009: User information made public without
permission
2014: manipulated news feed to see if the
system could assess mood
2018: Revealed Cambridge Analytica sold FB
data
115

116. Prediction Drawbacks
AI algorithms rely on past behavior to predict future
behavior
Programming and test set must define “normal” for the
system to detect “abnormal”
Cannot predictwhat has not already occurred
• Taleb’s black swans
• Flash Crash of 2009
Past behavior predictionignores present environment and
emotional influences
Must define normal to program for abnormal detection
116

117. Privacy Paradox
Privacy risk is weighed against value of object,
interaction, end result
• Research assumes user calculates an internalized value
• Basis for choice to reveal personal identification
information (PII)
Value is determined by the smoothness of the
interaction (Groupon, Amazon Local)
• Value proposition overrides security/privacy concerns
Higher level of user control over PII reduces the
perception of risk
117

118. Tim Cook on Privacy
Called on US to pass comprehensive data security act
along the lines of GDPR
4 guiding Principles
• Right to have personal data minimized
• Right to know what is being collected and why
• Right to data security
• Right to access
118

119. If you’re not paying for it, YOU are the product
119

120. AI Ethics
120

121. AI Ethics = TL2
121
Popular search engine returns 215,000,000
results for AI ethics

122. “Algorithms are opinions
embedded in code.”
Cathy O’Neill
Weapons of Math Destruction
(2016)
122

123. Algorithmic Bias
Technology inherits ideas and values of the group that
develops it
Algorithm development rests on emotional capitalism
• Emotional capitalism: feeling can be managed rationally and
governed by logic
• Emotional socialism: suffering is unavoidable and should be
tolerated
Accept decisions from an automated system as agnostic
3 types
• Implicit (absorbed automatically
• Accidental (introduced by ignorance
• Deliberate
123

124. Algorithmic Bias
Technology inherits ideas and values of the
group that develops it
Algorithm development rests on emotional
capitalism
Accept decisions from an automated system as
agnostic
124

125. Governance Issues
Explanation (transparency)
• Core components
• Local Explanation: explain for specific decision, not system as
a whole
• Counterfactual Faithfulness: expect the explanation to be
causal and can be provided without providing contents of
the system
• Provide in situations where a person would be required to
do so
Regulation
• Regulators don’t understand what they are regulating
• Risk of stifling innovation
Applications (consistency)
• Impact beyond decision-maker
• Know if AI behaving erroneously
125

126. Accountability Under Law
Explanation
• Core components
• Local Explanation: explain for specific decision, not
system as a whole
• Counterfactual Faithfulness: expect the explanation to be
causal and can be provided without providing contents
of the system
• Provide in situations where a person would be required
to do so
Regulation
• Regulators don’t understand what they are regulating
• Risk of stifling innovation
Consistency of Application
• Impact beyond decision-maker
• Know if AI behaving erroneously
126

127. Bias Remedies
Design thinking
HCI heuristics as well as performance benchmarks
HCI professionals testing prior to live site deployment
Diversity/bias audits
Accountability
127

128. 128

129. Explainable AI (xAI)
xAI = field of research addressing interpretability and
explain-ability in ML and AI
• Compliance with relevant legislation
• Broader range of debugging
• Those working on system learn from it
• Enhanced trust in system decision-making (including scenarios where it
can break down
AI is a black box for those outside of computer science
AI development must shift from ad-hoc models for
decision-making that is more trustworthy
• Contrastive (present alternative data points)
• Counter-factual (changes in features that would lead to a different
outcome)
129

130. UK Parliament AI CoC
130
Application of a cross-sector code for the development
of AI applications
• Artificial intelligence should be developed for the common
good and benefit of humanity.
• Artificial intelligence should operate on principles of
intelligibilityand fairness.
• Artificial intelligence should not be used to diminish the data
rightsor privacy of individuals, families or communities.
• All citizens should have the right to be educated to enable
them to flourish mentally, emotionally and economically
alongside artificial intelligence.
• The autonomouspower to hurt, destroy or deceive human
beings should never be vested in artificial intelligence.

131. AI NOW Initiative (2018)
Kate Crawford (Microsoft) and
Meredith Whittaker(Google)
Founded to deal with issues of AI
diversity and inclusion
Conduct empirical studies
focused on
• Bias and inclusions
• Labor and automation
• Infrastructure and Safety
• Basic rights and liberties
131

132. AI Safety
132

133. Not Good AI
Used in for negative
outcomes
• Autonomous weapons
• Biased facial recognition
Used for malicious
purposes
• Fake news
• Denial of attack
133

134. Generative Adversarial Networks
Dueling neural networks
• 1 to generate an image from a data set
• 1 to determine if the image came from the data set
AI cop and counterfeiter game of cat and mouse
134

135. AI Risks
Mis-specified Objectives
Negative Side Effects that extend to wider application
Hacking: rewards, devices
Bad extrapolation of the real world
Poor training data
Privacy
Fairness
Abuse
Transparency
135

136. AI Risk Mitigations
Define impact regulator
• Future state
• Substitutes lower impact null actions
Train impact regulator
• Over many tasks
• Separate training parameters for task side effects
Penalize influence
• Use information-theoretic measures to capture agent’s
potential for information
• Penalize empowerment
Provide scalable oversight with multi-agent approach
136

137. AI Risk Mitigations 2
Use Objective functions to capture designer informal
intent
• No partially observed goals
• Concrete, not abstract rewards
• Deep correlation between tasks and functions
Feedback loops
• Model look ahead
• Reward capping
• Counter example resistance – combination of rewards
137

138. AI Risk Mitigations 3
Safe exploration
• Risk sensitive performance criteria
• Use demonstration
• Simulated exploration
Well defined models
• Train on multiple distributions
• Program for out-of-distribution situations
138

139. Break
https://affinelayer.com/pixsrv/

140. DESIGN FOR MACHINES & HUMANS
The real meets the artificial
140

141. Human-centered design has expanded from the
design of objects (industrial design) to the design of
experiences (encompassing interaction design, visual
design and the design of spaces). The next step will be
the design of system behavior; the design of
algorithms that determine the behavior of automated
intelligent systems
Harry West
CEO, Frog Design
141

142. Machines Users Are Different
Logic: exacting, context independent, conditional logic
Development: uses explicit rules to define possible
behaviors
• Heuristics
• Intuition derived from huge data sets
142

143. 143
Information
Architecture

144. Information Architecture and AI
Problem definition and structure
Connections
Proto-typicality (mental models)
Visual complexity (rely on text more than images)
144

145. Form IA and AI Strategies
Customer Empathy Framework
• Define the problem
• Formulate the solution
• Map the environment (customer journey)
Tools
• Personas (use cases)
• Problem statements
• Environment description (include systems and
processes)
• Success benchmark success (quantitative, qualitative)
145

146. Create Meaningful Structures
Site Structure
• Machine readable text
• Related content model
• Schema markup
Internal linking to reinforce context relationships
and discovery
146

147. Structured Data
Name the
components on the
page for the
machine user
147

148. Navigation for AI
148
Name object for cross system compatibility
Move toward the center of project
Create deliverables that bridge the logical world of
IAs and the physical world of implementers
Converse with other disciplines in language they
understand and employ

149. 149
User
Experience

150. How UX Professionals Defined UX?
A consequence of a user’s internal state,
(predispositions,expectation, needs,motivations,mood, etc.)
the characteristics of the designed system
(complexity,purpose,usability, functionality,etc.) and the
context (or the environment) within which the
interaction occurs (organization/social setting,
meaningfulnessof activity, voluntariness of use,etc.)
150

151. Key UX Data Points
Conversions
Unique Visitors
Bounce rate
Social Actions
Number of Pages/visited
Average time on page (exclude bounces)
Exit rate
151

152. Panda Algorithm Negative Signals
High % of deep content
Low amount of original content
High amount of ads or gratuitous images
Large quantity of boiler-plate text
Over-optimized (too many links)
High bounce rate
Low visit duration
Low CTR from Google search results
No/Low quality in-links
No/Low social mentions
152

153. Google Optimal Page Layout
153

154. Observed Self better than Quantified Self
154

155. Use a Different Pattern Library
Visitor search patterns: Use online tools to
uncover customer intent
Visitor behavior patterns: website analytics
Visitor conversion patterns: content to address
all stages of conversion funnel
Tools
• Search suggest scrapers
• SEO|Content Marketing software
• Webmaster and website analytics accounts
155

156. 156
Content

157. Content & Context Algorithms
Hypertext (HITS) Induced Topic Search
Hilltop
Topic Sensitive PageRank
Orion (2008)
Hummingbird
157

158. HITS (1997)
Hypertext Induced Topic Search
HITS is a related algorithm for Authority determination
HITS = PageRank + Topic Distillation
Unlike PR, query dependent
Somewhat recursive
158

159. Authorities & Hubs
159

160. Hilltop Algorithm (2001)
Topic segmentation algorithm = query dependent
Introduces concept of non-affiliated “expert
documents” to HITS
Quality of links more important than quantity of links
Segmentation of corpus into broad topics
Selection of authority sources within these topic areas
160

161. Latent Semantic Indexing
Using a ~<search term>
will initiate Google’s LSI
and produce a list of
results that contains
your original term as
well as documents that
the search engine
determines are relevant
to your query.
161

162. Topic-Sensitive PageRank (2002)
Context sensitive relevance ranking based on a set of
“vectors” and not just incoming links
Pre-query calculation of factors based on subset of
corpus
Context of term use in document
Context of term use in history of queries
Context of term use by user submitting query
Based on 16 top-level Open Directory categories
162

163. Orion Algorithm (2008)
Purchased by Google in
April 2006 for A LOT
of money
Results include expanded
text extracts from the
websites
Integrates results from
related concepts into
query results
163

164. Hummingbird: Entity detection
Comparison of search query to general
population search behavior around query
terms
Revises query and submits both to search index
• Confidence score
• Relationship threshold
• Adjacent context
• Floating context
• Results a consolidation of both queries
164

165. AI Content Components
Traditional IR (tf*idf)
Link analysis for Authority
Location on page
Query type
Content Qualities
• Uniqueness
• Authoritative
• Freshness
• Well Written
165

166. Transform Keywords Into Intelligence
Keywords are user queries
Queries represent user information needs and
satisfaction threshold
Keywords become intelligence
• Competitive: who is doing better
• Visibility:how do the search engines see my content
• Customer: how do targeted customers look for my
products and Services
Tools
• Search suggest scrapers
• Google Trends
• SEO Software (BrightEdge, SEMrush)
166

167. Establish Context
Context becomes what the system can measure
• Environmental features
• Interactions
• Ubiquitous computing
• Internet of things (IoT)
• Digital Assistants
Non-methodical approach that brings in
containment (social through local) interactions
• Adaptive/reactive interaction in situ
• Context as perceived and used by actor
167

168. Create and Curate Content
Entities Rule
Newspaper model
Opening paragraphs most important for
subject determination
Relational content model
168

169. User interest mapped to customer
journey and content type
169
Keywords
Customer Journey Phase AVG CAT Rank _Monthly Searches Average of Competitor Rank
Consideration 32 73,200 #DIV/0!
Used 30 47,690 #DIV/0!
Sale 35 21,120 #DIV/0!
Product Information 29 2,990 #DIV/0!
Competitor 68 1,150 #DIV/0!
Quick Answer 32 110 #DIV/0!
Rental 6 50 #DIV/0!
Parts 56 40 #DIV/0!
Reviews 40 #DIV/0!
Competitor Rental 55 10 #DIV/0!
Purchase 39 2,980 #DIV/0!
Sale 42 2,690 #DIV/0!
Used 38 260 #DIV/0!
Product Information 14 30 #DIV/0!
Awareness 12 2,820 #DIV/0!
Used 17 1,640 #DIV/0!
Competitor 880 #DIV/0!
Product Information 9 250 #DIV/0!
Quick Answer 30 #DIV/0!
Sale 7 20 #DIV/0!
Post Purchase 30 30 #DIV/0!
Product Information 30 20 #DIV/0!
Used 10 #DIV/0!
Grand Total 32 79,030 #DIV/0!

170. Map Semantic Connections
Semantic technology requires everything to be
associated to understand user activity
• Control layer
• Mapping (semantic) layer
• Device layer
Semantic analysis model
• Semantic layering
• Semantic mapping (Boiko IAS 2018)
• Semantic machine heterogeneity
Association between user behavior patters (customer
journey map)
170

171. Give Users What They Want
171
# of pages in directory
# of pages views for
each directory

172. Exercise: Develop AI Application for Crisis Help
Line
1. Choose a model
I. Base Models
I. Learning
II. Prediction: create actions to respond to learning
II. And a sub-models
I. Data analysis
II. User Identification
III. Behavior recognition
IV. Service construction, Service Provisioning
2. Define objective function
3. Train system by adjusting parameter’s (reward) to
maximize objective function
4. Test to evaluate accuracy and effectiveness of the
model
172

173. 173
Design

174. Data Driven Design
Without a person at (or near) the helm who
thoroughly understands the principles and elements
of Design, a company eventually runs out of reasons
for design decisions...
When a company is filled with engineers, it turns to
engineering to solve problems. Reduce each
decision to a simple logic problem. Remove all
subjectivity and just look at the data. Data in your
favor?...
And that data eventually becomes a crutch for every
decision, paralyzing the company and preventing it
from making any daring design decisions.
174

175. Generative Design
AKA Mutative Design, Parametric Design
Designer defines rules for algorithm
Algorithm generates variations using the predefined
rules
Algorithm filters the results based on design quality and
requirements
Designer chooses the best variants and polishes as
needed
System runs A|B tests for variant(s)
Test results are used to choose most effective design
175

176. Privacy By Design
Opt in / Opt out
User control over sharing – notifications, time limits
Command user attention for privacy decisions
176

177. Not Privacy by Design
177

178. Visual Complexity & Prototypicality
178

179. Google Page Layout
179

180. AI Design According Computer Science
Components
• Variables
• Domains (environment)
• Constraints (limits)
Goal of AI Design = satisfy constraints
Admissible heuristic: if it costs too much to reach
solution state then revise or reject
180

181. Design Thinking for Data Science 1
Reach out to the development staff
Embrace design thinking
Transform “my idea” into “our idea” with early stage
collaboration
181

182. Design Thinking for Data Science 2
Customer Empathy Stage
• Understand the problem solving
• Define the solution
• Map the environment (customer journey)
• Define the characteristics of a good solution (heuristics)
Outputs
• Personas (use cases)
• Problem statements
• Environment description (include systems and
processes)
• Benchmark success (quantitative, qualitative)
182

183. Design Thinking for Data Science 3
Go Broad, Go Deep Stage
Brainstorm solution ideas across silos
Diversify contributors
Post all artifacts and review as a group
Organize ideas into themes
Include “leap of faith” assumptions
Take the best and formulate a solution hypothesis
183

184. Design Thinking for Data Science 4
Rapid experimentation with Customers
Paper prototyping, sketches, storyboard
Build stable testing methodology into plan
Start small (project | testing) to achieve collective wins
184

185. Smart Tools & Platforms
Semantic image segmentation
Font recognition
Intelligent audience segmentation
185

186. Data Protection by Design
Design strategy for accountability
• Enforceable policy
• Demonstrated Compliance
Detect and address bias
Components
• High-level design goals
• Privacy enhancing technology (user controls)
• Sanitation of data
Principle of Accountability
Discrimination Aware Data Mining (DADM)
186

187. Privacy Design Recommendations
Different UI for different tasks
Opt in, not opt out
Build in alerts if system deviates from the norm
Clear explanation of system decision making methods
and reasoning workflow (xAI)
Government enforced standards of data collection
and control
187

188. 188

189. Algorithm-Based Design 1
Designer as art director, algorithm as apprentice
Determine “well designed” site for learning model
Create mood board for algorithm to deconstruct
Use algorithm for simple tasks
• Color match up
• Image assembly (movie poster app)
• Styling videos
• Extract usage patters from data sets
189

190. Algorithm-Based Design 2
Designer and Developer define the logic to consider
content, context and user data
AEM (behavior targeted UI)
Brightedge DataMind
Vox Media Homepage Generator
190

191. Machine Learning Design Process
Define learning problem
• Inputs
• Outputs
• Types of training data needed
Generate good data
• Completeness
• Accurate
• Consistent
• Timely
Sketch out user and data flow (decision trees)
Test assumptions against prototype
Start with simple mechanism and move to complex
191

192. Thank You
Marianne Sweeny
Principal
Daedalus Information Systems
sweeny48@uw.edu
@msweeny
192
Embrace, engage, define, direct

193. APPENDIX
A friendly drop to help you after…
193

194. Suggested Reading
• Algorithms to Live By; Brian Christian, Tom Griffiths
• Super Intelligence: Paths, Dangers, Strategies; Nick
Bostrum
• The Tides of Mind: Uncovering the Spectrum of
Consciousness; David Gelernter
• The Knowing Project; Michael Lewis
194

195. Twitter Resources
195
Rob Wortham @RobWortham
Frank Pasquale @FrankPasquale
Luke Robert Mason @LukeRobertMason
Garry Kasparov @Kasparov63
John C. Havens @johnchavens
Joanna Bryson @j2breve, @j2blather
Carol Smith @carologic
Sentiment/Emotion/AI @SentimentSymp
Elizabeth Churchill @xeeliz
Adam Coates @adampaulcoates
Richard @RichardSocher
Yann LeCun @ylecun
Kirk Borne @KirkDBorne
Right Relevance @rightrelevance
Machine Learning @ML_toparticles
Andrew Ng @AndrewYNg
Atsushi HASEGAWA @ahaseg
Eric Horvitz @erichorvitz
Sander Dieleman @sedielem ?
AI Now Institute @AINowInstitute
Oren Etzioni @etzioni
Jeff Dalton @JeffD
Peter Trainor @petetrainor
Rob McCargow @robmccargow
Kevin Slavin @slavin_fpo
Giles Colborne @gilescolborne
Lev Manovich @manovich
Luke Robert Mason @LukeRobertMason
Jana Eggers @jeggers
Dawn Anderson @dawnieando
Colin Eagan @ColinEags
Data Science Central @DataScienceCtrl
Right Relevance @rightrelevance
Machine Learning @ML_toparticles
Brenda Laurel @blaurel
Ian Soboroff @ian_soboroff
Phillip Hunter @designoutloud
Paul Dourish @dourish
Jason Alderman @justsomeguy
manovich @manovich
Dorian Taylor @doriantaylor
Kirk Borne @KirkDBorne
Tim Caynes · @timcaynes.

196. 196