2. 2
Recruiting Tools
Amazon recruiting tool shut down for
bias against women after it codified
discriminatory practices due to narrow
data sets
99-100% 65-79%93-98% 88-94%
Recognition AI
Calls for regulation on use of facial
recognition after consistently higher error
rates for darker-skinned and female faces
3. 3
Black Box Risk Assessments
10+yrs of model behavior but denied
parole due to high-risk assessment
Details on over 100 factors and weights
protected as commercially proprietary
Single, subjective question lowered risk
scores from an 8 (of 10) down to 1
4. 4
China Social Credit System
Ranks citizens’ behavior to
determine their social and credit
worthiness
1.4B people will have a score by
2020 which will impact their social and
economic rights
5. Myth: Shiny is better
Myth: There’s an unavoidable trade-off
between accuracy and interpretability/privacy
Myth: All we need is more data
Myth: It’s ok to transfer AI created for
non-critical tasks to high-stakes decisions
6. As creators of artificial
intelligence systems, we have a
duty to guide the development
and application of AI in ways that
fit our social values
Responsible AI
Accountability
Fairness
Public Trust
9. 9
We observe, collect adjacent data, and
make connections
We process the connections and to learn
and make informed, in-context decisions
We make tens of thousands of decisions
daily, most of which depend on
surrounding circumstances and context.
45
10. 10
AI must access and process a great deal
of contextual, connected information
• Learn from adjacent information
• Make and refine judgements
• Adjust to circumstances
The fastest, most reliable way to
manage data connections is with graph
technology45
12. 12
“AI is not all about Machine
Learning.
Context, structure, and
reasoning are necessary
ingredients, and Knowledge
Graphs and Linked Data are
key technologies for this.”
Wais Bashir
Managing Editor, Onyx Advisory
13.
14. 14
Graphs are built for relationships
– with relationships
Imbue individual entities with
connections as a fabric
Enriches data so it is more useful
16. 16
EMPLOYEE
name: Amy Peters
date_of_birth: 1984-03-01
employee_ID: 7875
COMPANY CITY
:HAS_CEO
start_date: 2008-01-02
:LOCATED_IN
start_date: 2008-01-02
Neo4j invented the property graph in 2002 using a napkin sketch –
the connected-data model that still works today.
We can process millions of data connections per second and
perform analytics on billions of nodes
20. $72.5 Billion Opioid Insurance Fraud per Year
In frauds rings drugs are improperly prescribed
by doctors and filled by cooperating pharmacists,
all of whom pocket illegal payments
20
21. 21
Graph algorithms reveal clusters of interactions in
large networks to detect communities for ML
Predicting fraud accurately requires extreme
insight into the relationships among entities
Predictive accuracy
22. Driverless Cars Must Be Foolproof
Tesla autonomous car tricked into changing
lanes with stickers
22
23. 23
Adjacent data helps widen and deepen the scope of
AI systems so they are more broadly applicable in
their environments
Situational awareness is crucial when context-based
learning and actions are part of AI systems
4
5
Situational flexibility
24. 24
Gaming the System
High-stakes criminals misrepresented and
manipulated input data to fly under the radar
Detecting evolutionary financial statement fraud
25. Reliability and Explainability
25
When data is stored as a graph, it’s easy to track
how it changes, who changes it and where it is used
For AI solutions to be viewed as reliable the
underlying data needs to be reliable
Example from Neo4j Risk Mgmt. Solutions
26. Past and Current Data Amplifies Bias
Data skewed by discrimination and demographics
creeps into policing, programs and sentencing
26
COMPAS Scores at Booking
27. 27
Fairness
Understanding our data can reveal bias inherit in
the information, in how it’s collected or in how it’s
used to train our models
Graphs adds contextual information to our ML data
and reveals relationships within data – which are
often better outcome predictors than raw data
“…data without context is
just organized information.”
Albert Einstein
28. 28
Human Interaction is Crucial
Boeing fails to incorporate pilot reactions into
737 Max auto-pilot system
29. 29
AI systems can be over-fitted to tight scenarios and
idealized situations that don’t account for the range
of human interactions
Graphs encapsulate the way we think about the
world, making it easier to incorporate human
responses and explain outcomes / processesRobustness
Trustworthiness
31. AI guidelines that promote societal
values
AI solutions will increase situational
appropriateness, tamper-proofing,
explainability and transparency
Faster adoption of AI solutions as
they become more trustworthy
33. 33
The inclusion and use of adjacent
information as context for AI will
become a standard
This will drive more reliable, accurate
and flexible AI solutions
34. “The idea is that graph networks are bigger than any
one machine-learning approach.
Graphs bring an ability to generalize about
structure that the individual neural nets don't have.”
35. 35
Implements machine learning in a graph environment
Native graph learning will move today’s AI from a rigid, black box approach
to extremely flexible, accurate and transparent models
Lets
users input
connected data
Learns while
preserving
transient states
Produces
outcomes in
graph format
Enables experts to
track and validate
AI decision paths
More accurate with
less data, learning
important features
36. 36
“Coders are the most
empowered laborers
that have ever existed.”
Anil Dash @anildash
Glitch CEO
Ethical technology activist
37. 37
Know & Track Data
(Graphs for data lineage)
De-Bias Data
(AI Fairness 360 toolkit)
Learn/Ask for Help
(Algorithmic Justice League)
Involve Domain Experts
(Predictors, data, success)
Planning &
Data Collection
Train & Model
Results to
Implementation
Add Relationships
(Graph features,
Counterfactual search)
Look at Model Exchanges
(ONNX, MAX)
Use Interpretable Models
Where You Can
(Prediction Lab at Duke)
Add Context to
AI Predictions & Heuristic AI
(Knowledge graphs)
Use Formal & Independent
Risk Assessments
(Checklists to committees)
Insist on Explanations in
High-Stakes Decisions
(Accurate, complete, faithful)
38. 38
Know & Track Data
(Graphs for data lineage)
De-Bias Data
(AI Fairness 360 toolkit)
Learn/Ask for Help
(Algorithmic Justice League)
Involve Domain Experts
(Predictors, data, success)
Planning &
Data Collection
github.com/IBM/AIF360
youtube.com/watch?v=Y0KA5U81w3U youtube.com/watch?v=Y0KA5U81w3U
ajlunited.org/
39. 39
Train & Model
Add Relationships
(Graph features,
Counterfactual search)
Look at Model Exchanges
(ONNX, MAX)
Use Interpretable Models
Where You Can
(Prediction Lab at Duke)
40. 40
Results to
Implementation
Add Context to
AI Predictions & Heuristic AI
(Knowledge graphs)
Use Formal & Independent
Risk Assessments
(Checklists to committees)
Insist on Explanations in
High-Stakes Decisions
(Accurate, complete, faithful)
ec.europa.eu/digital-single-market/en/
news/ethics-guidelines-trustworthy-ai
fujitsu.com/global/documents/about/res
ources/publications/fstj/archives/vol5
5-2/paper14.pdf
41. 41
Know & Track Data
(Graphs for data lineage)
De-Bias Data
(AI Fairness 360 toolkit)
Learn/Ask for Help
(Algorithmic Justice League)
Involve Domain Experts
(Predictors, data, success)
Planning &
Data Collection
Train & Model
Results to
Implementation
Add Relationships
(Graph features,
Counterfactual search)
Look at Model Exchanges
(ONNX, MAX)
Use Interpretable Models
Where You Can
(Prediction Lab at Duke)
Add Context to
AI Predictions & Heuristic AI
(Knowledge graphs)
Use Formal & Independent
Risk Assessments
(Checklists to committees)
Insist on Explanations in
High-Stakes Decisions
(Accurate, complete, faithful)
42. 42
“A lot of times, the failings
are not in AI. They're human
failings...
…if you’re not thinking about
the human problem, then AI
isn’t going to solve it for you.”
Vivienne Ming
Executive Chair & Co-Founder, Socos Labs