Defining the boundary for AI research in Intelligent Systems Dec 2021

Research Scholars’ Colloquium 2021 (RSC'21)
on Intelligent and Smart Systems
Research & Industry Symposium in honour of
Shri. F.C. Kohli, Father of Indian Software Industry
Organized by
SSN College of Engineering and supported by
ACM,CSI & IEEE CS Chennai Chapters
Defining Smart and Intelligent Systems Boundary
Dr. Parasuram Balasubramanian,
Founder & CEO, Theme Work Analytics,
Bangalore, India, 560041
Dec 3, 2021

Coverage
v Smart Vs Intelligent Systems
v Artificial Intelligence Definition
v AI Devices Classification
v AI Applications: present and Potential
v Cyber Physical Systems
v How far do machines Collaborate with humans ?
v Domains of concerns
v Survivorship Selection Bias
v Bias in Data, Model (Methodology) and Decisions
v Eliminating bias in AI Application Design
v AI vs Causality Determination
v AI Pitfalls
v The AI Dilemma
v The Boundary Setting Problem in AI
v Concept of Human Parity
v Measurement Criteria
v UNESCO Guidelines for ethical use of AI
v GOI ( MEITY) AI Focus
v Drawing the Boundary Line
v The AI Researcher’s Commitment
Ada Lovelace
Grace Hopper
Gladys West

Smart vs Intelligent Systems
Smart Systems
Ø Devices with embedded
software.
Ø Collect and Transfer
Data.
Ø May perform pre-
programmed Data
Processing functions.
Ø Resultant Action
Execution Control with
humans
Intelligent
Systems
(Future)
Ø Have self learning
capability.
Ø The algorithm can revise
its functionality based
on new knowledge
gained, much like a
humans..
Intelligent
Systems
(Current)
Ø Contain software with AI
functionality.
Ø Perform data processing,
analysis, inference and
decision making to a
predetermined limit.

The field of AI began at a 1956 workshop at Dartmouth College in USA [top]
attended by, from left, Oliver Selfridge, Nathaniel Rochester, Ray Solomonoff,
Marvin Minsky, an unidentified person, workshop organizer John McCarthy, and
Claude Shannon
AI Definition
"every aspect of learning or any other feature of
intelligence can in principle be so precisely described
that a machine can be made to simulate it."
Ø Intelligence demonstrated by machines.
Ø Machines performing tasks that require human
intelligence.
Ø Systems demonstrating human like intelligent
behaviours such as sensing, planning, learning,
reasoning, problem solving, knowledge
representation, perception, motion, and social
intelligence and creativity ( from European
Parliament Committee)
Ø IBM included Decision Making explicitly in its
definition.

The field of AI began at a 1956 workshop at Dartmouth College in USA [top]
attended by, from left, Oliver Selfridge, Nathaniel Rochester, Ray Solomonoff,
Marvin Minsky, an unidentified person, workshop organizer John McCarthy, and
Claude Shannon
AI Definition
Ø Many Auxiliary Questions
Ø Is it enough to teach machines to learn?
Ø Can AI machines be easily fooled?
Ø Can they ever learn how to handle new situations?
Ø How can we program them for non-interference, when
multiple robots are present in one scenario?
Ø Is it possible to teach the machines to deal with nuances
and caveats?
Ø Can a robot be programmed from being switched off?
Ø Can a robot be endowed with legal rights?
Two Fundamental Questions
Ø What tasks can be assigned to the machine?
Ø Can it be allowed to control execution?

AI Devices Classification
Reactive Machines Limited Memory Theory of Mind Self Awareness
Reference: https://www.govtech.com/computing/understanding-the-four-types-of-artificial-intelligence.html November 14, 2016 • Arend Hintze, Michigan State University
§ First Gen Robots,
Expert systems
§ No Memory
§ Cannot use past
experience in
current decisions
§ Run on algorithms
§ Need to be
programmed for
specific tasks.
§ Considered as
simple AI systems
§ Autonomous Cars
§ Use limited
memory to store
recent past data
to compare with
current data.
§ Transient memory,
§ Collaborative
Cyber Physical
Systems
§ Can understand
and mimic human
feelings and
thoughts.
§ Can modify
behaviour to
other objects’
expected
behaviour in the
neighbourhood
§ Currently in
science fiction
§ Machine with a
conscience.
§ Futuristic.
§ Self learning .
§ Can feel, think
and act like
humans.
Current
boundary

Alan Turing John McCarthy Marvin Minsky Ray Kurzweil
AI Applications : Present & Future
Manufacturing
Hazardous Area measurement
Asset Maintenance
Life cycle tracking of manufactured goods
Collaborative Production
Health Care
Drug Discovery
Genome Sequencing
Cancer Detection, Medical Diagnosis
Personalized Medicine
Assisted Care for elderly
Bio Informatics
Financial Services
Fraud Detection
Credit Risk Analysis
Portfolio Management
Government
Services
Traffic Management
Border Protection
River Management
Education Sector
Machine Translation
Lesson Tailoring
Agriculture
Crop & Soil Monitoring
Weather Prediction
Fertiliser & Pesticide Spraying
Internet and
e Commerce
Search Engine
Recommendation Systems
Natural Language Processing
Transportation
Autonomous Vehicles
Multi modal Design
Safety Management
Media & Entertainment
Media Programming
News Filtering
Story Generation
Security & Law
Enforcement
Person Identification
Image Recognition
Terrain Mapping
AI Core Technologies
v Machine Vision
v Speech Recognition
v Natural Language Processing
v Robotics
v Machine Learning
Sanghamitra Bandyopadhyay

AlphaFold Is The Most Important Achievement In AI—Ever
https://www.forbes.com/sites/robtoews/2021/10/03/alphafold-is-the-most-important-achievement-in-ai-ever/
Rob Toews forbes Oct 3, 2021,07:34pm EDT
DeepMind's AlphaFold
represents the first
time a significant
scientific problem has
been solved by AI
what has AI actually accomplished or enabled that makes
a difference in the real world?
This summer, DeepMind delivered the strongest answer
yet to that question in the decades-long history of AI
research: AlphaFold, a software platform that will
revolutionize our understanding of biology
One of Life’s Great Mysteries
In 1972, in his acceptance speech for the Nobel Prize in
Chemistry, Christian Anfinsen made a historic prediction:
it should in principle be possible to determine a protein’s
three-dimensional shape based solely on the one-
dimensional string of molecules that comprise it.
Finding a solution to this puzzle, known as the “protein folding problem,”
has stood as a grand challenge in the field of biology for half a century
AI Applications : Present & Future

Raj Reddy
Cyber Physical Systems
Robotics
Robotics Drones
Cobots
Ø How to ensure Human and Equipment Safety in the vicinity?
Ø How to ensure Collaborative Work ?
Ø How to ensure Safety in Collaborative Work?
Ø How to ensure Privacy?
Jitendra Malik
Joseph Engelberger
v Subordination, Collaboration & Command
v Data Sharing vs. Decision Support
“ the obligation of
machines is to try to
optimize that aggregate
quality of human
experience”. [Customer
centric goals rather than
Seller focused goals]

How far do machines collaborate with humans?
Image Credit: https://www.cas.org/
It is harder to compete than collaborate with
human beings.
A 2021 Oct study at MIT Lincoln Laboratory
found that humans find it difficult to
understand the machine moves, feel they are
opaque and random : hence don’t trust
them.
It becomes an issue while defending missile
attacks or performing complex surgery,
together.
The field of teaming Intelligence (TI) is still
evolving

Domains of Concerns
v Human Safety
v Property Protection
v Information Security
v Bias in Data, Model( Methodology)
and Decisions
v Data Comprehensiveness in
representation
v Human Comprehension
v Absence of Causality
v Fear of the Unknown
Sunita Sarawagi

Survivorship Selection Bias
During WW2 bomber planes were returning to base in USA ,
after a combat with the German air defense guns.
They were examined to identify most hit areas , so that
protective armour can be added in select areas.
Abraham Wald, A Statistician from Columbia University
recommended the opposite ; that the non hit areas of cock
pit and motor need to be protected.
His argument that planes hit in those areas did not survive
or return. Hence they need to be strengthened.
That is known as Survivorship Selection Bias.
A data set is incomplete, if it contains only those that went
past successfully a selection criteria that is biased.
By Martin Grandjean (vector), McGeddon (picture), Cameron Moll (concept) - Own work,
CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=102017718

Bias in Data, Model (Methodology) and Decisions
Data Bias
Ø Reporting Bias
Ø Sample Bias *
Model Bias
Ø Mistaking Correlation
to causality
Ø Untested assumption
in Hypothesis
(Confirmation Bias)
Ø Confounding Variables
(lack of knowledge)
Ø (Over or under fitting)
Decision Bias
Ø Group Attribution
Ø Societal Bias
Truth Belief
Evidence
Accepted
Evidence
not used
Bias of any kind results in
incorrect decision.

Eliminating bias in AI Application Design
Data Bias
Ø Homogeneity check
Ø Sample Adequacy Check
Ø Proportional Representation check
Ø Pre checking intended output
parameters of the study
Model (Algorithm) Bias
Ø Process transparency
Ø Output explain ability
Ø Training data availability
Ø Algorithm steps being monitored
for intermediate outputs
Ø Building human in the loop for decision veto
Ø due consideration of such circumstances
pre- defined
Ø Building options at many steps
Ø Using a diverse team
Ø Correlation Causality trap avoidance

AI vs Causality Determination
ØML predicts outcomes but don’t understand Causality.
ØIt detects correlation but not causation.
ØLacks Generalization capability.
ØKnowledge of Cause Effect relationship is needed to
discover better solutions, Do What if Analysis without
material loss and to be cost effective
For many business applications such as Price Prediction, Object
Classification, Market Segmentation correlation is good enough.
But not for drug discovery, secondary effects of collision in AVs and
Robots
Reinforcement Learning, Hybrid
models of DL and Symbolic
Logic are WIP to evolve Causal
AI. But miles away.
Josh Tenenbaum, a
professor at MIT’s Center
for Brains Minds &
Machines and team have
designed an experiment to
show how DL falls short in
causal analysis

AI Pitfalls
Recommendation Engines in YouTube, Facebook etc. pushing
extreme or inflammatory content; indirectly assisting polarization
Gender, Race bias detected in Recruitment, Default Prediction and
Criminality assessment, Loan renewal
Unconscious bias in Society ( Stereotyping) unwittingly carried into
the model; remains undetected.
Unanticipated accidents, failures and loss in Autonomous Vehicles,
Cobots and Drones.
Over or under prediction of treatment levels, efficacy impact etc. in
health sector ( cancer detection, drug discovery)
Image manipulation possible through Adobe Photoshop’s super
resolution feature. Chrome, Facebook and Alexa seem to be
invading into our private space uninvited.
Technology in the hands of
the Darth Vaders?
Can we dare to put the code
as open source?

Data Gathering
Data Sharing
Decision Support
Decision Making
Decision Execution
Human
Safety
Privacy Causality
(Unknown)
Ethical
Security
Bias
Drones in
beach
AI for Riot
Control
Nursing Care in
emergency
Ward
Donee
Selection in
organ
transplant
Collateral
damage in AV
collision
Lesson tailoring
data with
Recruitment
System
Recommendation
Algorithm for
Cross Selling
(Target:Pregnancy)
Facial
Recognition
for Person
Authentication
AI for Loan
Default
Genie out of
the bottle

The AI Dilemma
The R2D2
Camp
Welcome AI with open arms. It can lead to unprecedented levels of
Productivity and Societal Advancement. It will be subservient and we
can coexist Harmoniously.
The
Terminator
Camp
AI is a minefield with potential for vast destruction as we walk
through. AI will dominate and control humans. We will be enslaved. It
is the biggest threat to mankind.
Ø Can we set the boundary in between Decision Support &
Decision Making instead of Decision Execution?
Ø Then we can retain the advantages of the AI Technology and still
create/preserve employment; Keep the world safe.
“ AI developments are more profound than fire or electricity” says Sundar Pichai.
Or Are we “summoning the demon” as Elon Musk thinks?
OR

The Boundary Setting Problem in AI
Ø Any research problem requires to be bounded. It has to define its scope and coverage.
Ø Universities and Research labs set these boundaries adhering to the legal framework and
societal considerations.
We run into
v Ethical issues when dealing with health care research
v Moral dilemmas in Genetic work
v Livelihood versus environmental protection in manufacturing industry
v Human Safety consideration in robotics
AI research and application go far beyond these issues; because
Ø What would confront us beyond the border scares humanity; It is an irreversible journey .
Ø The boundary needs constant resetting based on advancement and knowledge.
Ø It is a complex region with many factors under consideration.
Ø It pits scientists against policy makers.

. India Bangladesh Border prior to 2015
Ø Landmass of each in other’s territory
Ø 173 Enclaves in total
Ø Enclaves within Enclaves
Ø Weird problems of needing passport and visa
Ø More than 50,000 residents
Ø Exchange Territories
Ø Allow people freedom to move over, change or
retain citizenship
Ø Facilitate value realization
Problem
Solution

Concept of Human Parity
AI field Image/Object
Recognition
Speech
Recognition
Text reading
comprehension
Machine
Translation
Year Achieved 2016 2017 2018 2018
Accuracy
level (MS Data)
96% Error rate
5.1%
88.493% 69.9%
Machine Performance at an equal level of trained human experts
Good Enough to let
the machine handle
the task?
Ø It depends on the task and the application domain.
Ø For Security Clearance based on facial recognition, may not be sufficient.
Ø For delivering drugs based on care giver’s voice command, may be dangerous.
Ø For translation of sensitive information on an event ,to be shared in mass media, it may fall short.
Ø Can our “ value of life” considerations be handled fairly by machines in Organ Transplant?
Should we not hold the machine to a higher level of
responsibility than humans, in case of accidents?

Measurement Criteria
ØHuman Parity
ØAbsolute Error rate within 3 to 6 sigma.
ØEconomic Loss
ØInsurance Premia
ØSample Size needed
ØTrial Period until First Incidence
ØAbility to perform What If Analysis of the Context
Photo licensed under CC BY-NC-SA
Insurance Premia combines the loss value with loss probability. Sample Size is for destructive testing where
needed. Trial Period is used when we are unable to foresee the bearable risks. What If analysis of the Context is
required when Causality is yet to be determined.

UNESCO Guidelines for ethical use of AI
July2,2021 Paris Agreement among Member States. ;
to be ratified in November Conference:
Focused on promoting Human Rights and
Sustainable Development Goals
Issues covering transparency, Privacy, data
management and accountability covered in multiple
domains. Not to replicate real world biases online.
To provide governments and policy makers with a
global framework for AI. Regulation
Do no harm; to be ensured through risk assessment
procedures apriori
AI methods chosen should be appropriate to the
context and based on rigorous scientific foundation
Safety risk ( unwanted harm) and Security risk (
system vulnerabilities) addressed through entire life
cycle
Fairness and non discrimination across countries
adhering to the International Law.
Privacy, Data Protection adequacy ensured through
policies, guidelines and enforcement mechanism.
Human oversight on all critical decision making.
Ultimate responsibility and accountability to rest
with humans all the time.

GOI ( MEITY) AI Focus
Ø Playing a proactive role to promote and to regulate AI
Developments.
Ø Expert teams have been formed to recommend the Form and
Structure and Rules of Engagement with AI.
Ø Has created a National AI portal for Knowledge Sharing.
Ø Published reports on Application Priority Domains for the Nation,
Data and Platform Management, Skill Development and Policy
Guidelines and structure for addressing concerns of Cyber Security,
Safety and Ethical issues.
Ø More work remains to be done.

Drawing the Boundary Line
It has to deal with
Ø Tasks Delegated or Assigned to the AI Device
Ø Domains of concern
Ø Classification of Core Vs Application of AI Technology
Ø Business Application areas
Specify the role and distinguish the accountability of
v Government ( Central or State),
v Industry Body and the
v Firm/ Institution
v Researcher

Drawing the Boundary Line : To act with Expediency..
Ø Form a Cyber Physical Systems Commission (CPSC) at the national
level
Ø Empower CPSC to issue Domain Specific Guidelines for all CPS ; to form
Special Focus Groups that are Application Specific (SFG-AS)
Ø Empower each SFG-AS to approve major products and solutions for
mass usage, through a comprehensive apriori testing procedure and
monitor field adherence.
Ø Legislate adherence to a self regulation process at the firm and
individual researcher level and ensure compliance.

Thank you
Drawing the Boundary Line : Researchers’ commitment
Ø To define research boundaries with abundant caution.
Ø To ensure data is unbiased
Ø To Alert against use of their research for unintended and
questionable purposes.
Ø To be mindful of need for causal analysis even in black box
techniques.
Ø To always perform a potential impact and consequence analysis of
their research recommendation
Many professions seek the performer to adhere to a code of conduct. Like the Hippocratic oath for Physicians
and Oath of Allegiance taken by the legislatures. AI researchers need to evolve and adhere to a personal and
voluntary Code of conduct.

Computer Science Pioneers
Ada Lovelace
19th
Century
Worked with Charles Babbage in his analytical Engine.
Considered as the world’s first computer programmer.
Grace Hopper
20th
Century
Computer programming pioneer. Machine independent Language originator. Lead to one the first high level
languages COBOL
Gladys West
Mid 20th
Century
Mathematician whose work lead to the invention of GPS
Alan Turing
(1912-1952)
Mathematician. Considered as the Father of theoretical computer science and AI
John McCarthy
(1927-2011)
Computer scientist, one of the AI pioneers.
Developer of LISP language and Time Sharing concepts.
Marvin Minsky
(1927-2016)
Cognitive and computer Scientist. AI Pioneer. Cofounded MIT’s AI laboratory. Laid the foundation for Artificial
Neural Networks.
Ray Kurzweil
(1948-)
Inventor and a Businessman. Student of Minsky. OCR, Text to Speech synthesis, Speech Recognition Technology
and more.
Sangamitra
Bandyopadhyay
Computer Scientist , Director at Indian Statistical Institute. Leader in Computational Biology, AI, Patter
Recognition, ML, Bio Informatics
Joseph Engelberger
(1925-2015)
Physicist, Engineer & Entrepreneur. Considered father of Robotics. Invented the first industrial robot Unimate in
1950s.
Raj Reddy
(1937-)
Computer Scientist. Leader in Robotics and AI. Founder the Robotics Institute at Carnegie Mellon University.
Jitendra Malik
(1960-)
Computer Scientist And Professor at UC, Berkley.
Computer Vision Expert.
Sunita Sarawagi Computer Scientist and Professor at IIT Bombay. Leader and Expert in databases, data mining, ML, NLP.
Josh Tenenbaum Professor of computer Science at MIT. Computational Cognitive Science Expert. Mathematical Psychologist.

Cyber Physical Systems: role of Digital Twins
Ø Digital Twins Technology works by creating a Digital Image of
the Asset as well as by preserving the domain knowledge
relating to the asset’s technical operations in a given
environment.
Ø Hence it facilitates What If analysis performed on the DT to
evaluate scenarios ahead of time to guide in optimal decision
making. The best decision can then be operationalized on the
Physical Asset.
Ø IoT data collected in the field, can be fed to the DT to ascertain
the true and current status of the equipment and also to
record the results of Action taken.
Ø The latter can be utilized to validate the predicted and expected
results of the action (as determined earlier in the What If
analysis} and the correction needed for the predictive
algorithm.
Ø Hence it is a superb value adding tool to enhance human safety
and minimize cost of design or operations.
Ø The simulation and optimization capabilities of the DT form part
of its AI features.
Ø DT can be a means to determine
the boundary for safe operations, in
complex systems.

Conceive & Design
Implement & Maintain
Provide Services
HRD for Smart Systems: Skill Sets Trifurcation Model
Generic
Products and
Services
Apply at
Specific
Client sites
To end
customers
Design Skills
Impl & Maint. Skills
Deployment Skills

USA Jobs at High Risks over two decades
MostVulnerable
Retail Salesperson
Fast food and counter workers
Secretaries & Admin Assistants
Cashiers
Office Clerks
15 million jobs would be lost
by 2025
LeastVulnerable
Registered Nurses
School Teachers
General Managers
Software Developers
First Line Supervisors
Routine physical and cognitive function tasks face the highest risk of
elimination. Jobs with substantive human interaction, coordination and
tech development are projected to grow.

“Jobs Lost, Jobs Gained : Workforce Transitions in a time of Automation “ Mckinsey Global Institute report Dec 2017
Under midpoint scenario for automation
adoption 2016-30 Jobs Lost, Jobs modified
Global
China India
USA
Germany
15 % of
work force
= 400m
24 %
9 %
16 %
23 %
TransitioningWork Force =75m to 375 m
50m
60 % of the jobs have automation possibility of atleast 30 % of the tasks
long term scenario….
The Mckinsey study with a wide scope estimates the loss
of nearly 400 m jobs globally by 2030

Defining the boundary for AI research in Intelligent Systems Dec 2021

More Related Content

What's hot (19)

Similar to Defining the boundary for AI research in Intelligent Systems Dec 2021 (20)

More from Parasuram Balasubramanian (20)

Recently uploaded (20)

Defining the boundary for AI research in Intelligent Systems Dec 2021