Understanding the

Changing AI Environments
and The Need to Adapt
Herman Tan
Visiting Research Fellow
Institute of Operations Research and Analytics
Smart Nation Research Cluster
Singapore Experience
• Smart Nation
• Smart Business
• Safe AI
Smart Nation
• Initiative announced 24 Nov 2014
• Digital Transformation of Singapore
• Organized under Prime Minister’s Office
• Smart Nation and Digital Government Group
• Government Technology Agency
National Projects
• Codex – digital platform for development of govt apps
• E-payments – frictionless payment
• LifeSG – personalized government services and support
• National Digital Identity – Link all personal info in one app
• Smart Sensor Platform – National IOT network
• Smart Urban Mobility – Autonomous Vehicles
• GoBusiness
• Smart Sustainable Punggol
E-payments SG QR
NDI - SingPass
Personalised access to government services
Smart Nation Sensor Platform
• Smart Lamp Posts
• Wireless sensor network
• Drowning detection in pools
• Personal alert buttons for elderly
Smart Urban Mobility
National AI Office
• Under Smart Nation Initiative
• National AI Strategy

By 2030, Singapore will be a leader in developing and deploying

scalable, impactful AI solutions, in key sectors of high value and
relevance to our citizens and businesses

AI is the capability to simulate intelligent, human-like behaviour in

computers
Intelligent freight planning system
Seamless & efficient municipal services
Chronic disease prediction & management
Personalised education through adaptive
learning & assessment
Border clearance operations
AI Ecosystem Enablers

3
Talent Data
Helix

Trusted Environment
Smart Business
Growing Smart Businesses
• Global demand in new products
and services as solutions
• AI as a key enabler
• Government provide the
ground work and incentives
Adoption in Business
AI Ecosystem
• AI Singapore
• Public education – AI for everyone
• Build talent – AI Apprenticeship
• Industry projects – 100 Experiments
• Data Science Consortium
• Try AI for 1 year for free
• Institute of Data Science
• Research collaboration projects with Industry
• Institute of Operations Research and Analytics
• PhD program
• Industrial collaboration projects for PhD students
Adoption of AI in Business
• Supply of talent
• Data Science Consortium – try for 1 year
• Training existing talent
• AI Singapore – AI apprenticeship program
• Building demand for AI
• Government led grand projects
• Subsidies and grants for businesses to create value with AI
Existing Businesses
• How to leverage on emerging technologies to build value?
• Increasing revenue
• Reducing costs
• Reducing cycle time
Maturity of Analytics Capabilities

timoelliott.com
AI used in Manufacturing
New Start-ups
• Global market
• Emerging technologies
• Cloud native deployments
• Cloud-native solutions for global users
Singapore Innovation Hub
• Enterprise SG – Startup SG
• Centers of Innovation
• JTC Block 77
• NUS Enterprise
• NTUituve
• SMU – innovation academy
• A*Star
• AI Singapore
Safe AI
AI For Good
New Art
AI Fails
March 2019 – 51 killed by radicalized shooter
Youtube, Facebook, and other social media platforms were
instrumental in radicalizing the terrorist who killed 51 worshippers in a
March 2019 attack on two New Zealand mosques, according to a new
report from the country’s government.

WIRED.COM CECILIA D'ANASTASIO 12.08.2020

27 Mar 2018 – Tesla Crash

AFP NEWS AGENCY 1 April 2018

22 July 2015 – Robots used in surgery has
killed 144 since 2000

https://incidentdatabase.ai/
Safe AI Framework in Singapore
• Decisions made by AI should be Explainable, Transparent and Fair
• AI Systems should be Human-Centric
Brief History of Machine Learning
Customer Understandable Model
Factor 1 (F1) Payment history
Factor 2 (F2) Amounts owed
Factor 3 (F3) Length of history
Factor 4 (F4) New credit
Factor 5 (F5) Types of credit used

Probability of Approval = a x F1 + b x F2 + c x F3 + d x F4 + e x F5
Level of importance 35% 30% 15% 10% 10%
Correlation Plot between factors
Short History of Deep Learning
• 1989 – Yann LeCun developed LeNet for recognizing digits
• 2009 – ImageNet started
• 2010 – ImageNet competition started
• 2012 – Alexnet by Jeff Hinton won competition by big margin
 Image Net Winning Algorithm Performance
Bad CNN

Good
Deep Learning Success was a Surprise
• Neural net idea from 1960s
• Faded away as it could not achieve what it promised in 1990s
• 1998 Yann LeCun – LeNet went unnoticed
LeNet - 1989
Difference between Traditional ML and DL
More neurons → More accurate

Millions of neurons
More neurons → More accurate
ImageNet
ImageNet – Ontological Image Training Dataset
We can interrogate human decision
Can we interrogate an algorithm?
Deep Learning Model
Image Classification
Classification of tumor in X-Ray
Deep Learning Neural Network

Result

100s of millions of parameters

Safe AI
• Decisions made by AI should be Explainable, Transparent and Fair
• AI Systems should be Human-Centric
Explainable, Transparent and Fair
Customer Understandable Model
Factor 1 (F1) Payment history
Factor 2 (F2) Amounts owed
Factor 3 (F3) Length of history
Factor 4 (F4) New credit
Factor 5 (F5) Types of credit used

Probability of Approval = a x F1 + b x F2 + c x F3 + d x F4 + e x F5
Level of importance 35% 30% 15% 10% 10%
Accuracy vs Explanability
• 5 Factors used to approve loans Loan Approval Form
• Name
• Address – Postal code
• Payment history
F1 F2 F3 F4 F5 • Amounts owed
• Length of history
• New credit
• Types of credit
a b c d e

YES
Accuracy vs Explanability
• Factors that improve accuracy of loan default significantly
Loan Approval Form
• Gender, Marital status, Religion, Race
• Name
• Address – Postal code
F1 F2 F3 F4 F5 Name Address • Payment history
• Amounts owed
• Length of history
• New credit
a b c d e f g • Types of credit

YES

• Can you use this in the model for loan approval?

Accuracy vs Explanability
• Accuracy can be improved using a deep neural network
• Cannot be explained in a formula or showing coefficients Loan Approval Form
• Name
• Address – Postal code
F1 F2 F3 F4 F5 • Payment history
• Amounts owed
• Length of history
• New credit
• Types of credit

NO Why not?

• Can you use this in the model for loan approval?

Accuracy vs Explanability
• Can you approve using standard model but offer different terms to
different customers?
• Male just recently separated – 15% interest
• Female just married 2 years – 6% interest
• Religion = X gets 7% interest vs Religion = Y gets 9%
• Indians gets 9% interest vs Chinese gets 6% interest
Explainable Model used in Defect Diagnostic
Human Centric
Governed by humans
• Governance Council
• Chief Data Officer and Chief Privacy Officer
• Chief Information Security Officer
• Data Science Teams
• Risk assessment of the AI activity
• Alignment with corporate objectives
• Data sources utilized and potential biases
• Impact on individuals from AI decisions
Severity and Probability of Harm
• Shortest route recommendation
• AI Medical Diagnosis
Where should humans be included?
Example
Mastercard is a technology company in the global payments industry.
Its global payments processing network connects consumers, financial
institutions, merchants, governments and businesses in more than 210
countries and territories.
To achieve its vision, Mastercard leveraged AI in many applications such
as
• Fraud prevention
• Forecasting future spending trends
• Improving user retail experience.
What can go wrong?
AI/ML Operations
Bias in data

Omission bias Stereotype bias

Bias detected by Model Explainability
Google – “clownfish”
Deep Learning Neural Network

Intermediate results

Result

100s of millions of parameters

Image Classification
Classification of tumor in X-Ray
Misclassification of humans as primates

?
man

chimpanzee
Misclassification of humans as primates

?
man

toddler

chimpanzee
Earn Public Trust
• General disclosure
• Policy for explanation
• Explanability and transparency
• Interacting with consumers
• Option to opt-out
• Communication channels
• Testing user interface
• Acceptable user policies
• Interacting with other organisations
• Ethical evaluation
Self Driving Car
 Moral Machine Experiment

Nature, 1 Nov 2018 Global Summary

Preference differs by region
Where should humans be included?
Tesla
Where should humans be included?
• Smart Nation
• Smart Business
• Safe AI