This document discusses brainstorming ideas for experimentation approaches in AI/ML. It covers various topics such as the vision and mission for using AI, challenges and opportunities of AI, different types of human and machine reasoning, biases and fairness in AI, how to conceive experimentation ideas, how to onboard AI into practice, different types of data features, visualization methods, statistical and machine learning methodological approaches, and how XOPs can bridge humans and AI to build a better future.
7. Comparison of AI guidelines – Trigger for Exploration
Values to be
respected
AI Utilization
Guidelines
Draft AI R&D
guidelines for
international
discussions
Social Principles of
Human-centric AI
Ethics Guideline
for Trustworthy AI
Recommendati
on of the
Council on
Artificial
Intelligence
Ethically Aligned
Design
Asilomar AI
Principles
Tenets
by The Conference
toward AI
Network Society
(MIC) /Japan
The
Conference
toward AI
Network
Society
(MIC) /Japan
Integrated Innovation
Strategy Promotion
Council.(Social
Principles of
Human-centric AI)
/Japan
European
Commission (High
Level Expert
Group on AI (AI
HLEG))
OECD IEEE Global Initiative on
Ethics of Autonomous
and Intelligent Systems
Future of Life
Institute (FLI)
Partnership on AI
Human Centered
Human dignity
Diversity and Inclusiveness
Sustainable society
International Cooperation
Proper Utilization
Education/literacy
Human intervention and Controllability
Proper data
Collaboration among AI systems
Safety
Security
Privacy
Fairness, equity, removal of Discrimination
Transparency, Explainability
Accountability
…
https://www.soumu.go.jp/main_content/000637845.pdf
8. BIAS
• Simpson
Paradox
• Longitudinal
Data Fallacy
• Behavioural
• Content
Production
Linking
• Temporal
• Popularity
• Algorithmic
• User Interaction
• Presentation
• Ranking
• Social
• Emergent
• Self-Selection
• Omitted
Variable
• Cause-Effect
• Observer
• Funding
FAIRNESS
• Direct
• Indirect
• Systemic
• Statistical
• Explainable
• Unexplainable
• Equalized Odds
• Equal
Opportunity
• Demographic
Parity
• Through
Awareness
• Through
Unawareness
• Treatment
Equality
• Test
• Counterfactual
• Relational
Domains
• Conditional
Statistical Parity
• Individual
• Group
• SubGroup
govern
https://arxiv.org/pdf/1908.09635.pdf
Survey on AI Bias and Fairness – Trigger for Experimentation
10. learning from data ..
lifecycles of “above”..
sell gpu..
adopt
Op’s
ai
structured
image, text
unstructured
GPU’s
( DL )
data is both the representation of
worlds problem and its inherent
solutions ..
approach.. learning from data.. evolution
Data - Key Influence Solutions
data algorithms architecture
12. human reasoning…
• similarities patterns and make associations ( neighbour, grouping )
Analogical
• involves reasoning from a specific case or cases to derive a general rule ( trees )
Inductive
• making best guesses, dealing with uncertainties ( prior, naive)
Abductive
• independent analysis of parts (trustworthiness )
Decompositional
• linkage between two events (interpretability, explainability)
Cause-and-effect
• definitive ( data, meta data of different stages )
Critical thinking
• logical certainty (checks and bounds , context )
Deductive
15. How to conceive experimentation ideas.. ?
Brainstorming
•Metaphors
•Mockery
•Problem
•Failures
•Survey
Technology
Curiosity
•Stats vs ML VS DL
•DL & Decision
Trees
•Disassociate
Implicit Steps –
SVM DT
•Splitting Units
Question Terms
and
Terminologies
•Ground Truth
•Cross Validation
•Weaker Signatures
•Boosting
•Bagging
•Dropout
Inspiration
from Failures
•Stability in Human
Recognition
•Which Stride,
Which Convolution
•Which is more Fair
?
Alarming
Failures
•Cyber Security
National
Emergency
Progress
•Curiosity
•Specificity
•Finer
Understanding
16. Deployment
Model
Concern
Data
Knobs
for
Audit,
Regulations
Class
•Labelled ?
•Ranges ?
•Consistent
Correlation ?
•Where Mis-
Classification?
•Left out
Binary
•Fuzzier
Boundary
Multi-Label
•Multi –Class
•What more
• Framewor
k for Data
Governan
ce
Insights
Simplify
Which
Model
Splitting
Units
Hyper
parameter
Feedbacks
Framework
for Model
De-
Mystifying
Bounded,
Context,
Constructs
Generalization
Errors
Drift in Data,
Model,
Mathematical
• Framewor
k for
• Heuristic
• Extraction
Secure
Attack
Adversity
Framework
for AI
Testing
Weaker
Signatures
Boosting
Bagging
Dropout
Framework
for AI
Governance
Take 1
problem at
time.. to
explore .. the
dot’s
XOP’s based
approach
quantify
observation
How to progress on experimentation ideas.. ?
problem observation solutions Observability
18. cater 2 life … scenarios .. purpose ..
Making Sense
life
• Afghan<>Taliban
• Giving birth under the Taliban
• https://www.bbc.co.uk/newsround/24118241
Task
Nature
• Amazon Forest
• Task
Identify
19. How do you go about reading ..?
Data Engineer
Data
Analyst
Visualization
Story Telling
Statistician
Descriptive
Inferential
Data Scientist
Mining
XOPS
Continuous
https://www.bbc.co.uk/newsround/24118241
Driving
Factors
Purpose
Question
Evidences
Answers
20. Data Capture
Age No of
Child
Pain
Relief
Medic
ine
Food Temp Power Fuel
3 No No No 43C No No
https://www.bbc.co.uk/newsround/24118241
21. Data Capture
Age No of
Child
Pain
Relief
Medic
ine
Food Temp Power Fuel
3 No No No 43C No No
https://www.bbc.co.uk/newsround/24118241
Correlation? ( Association )
Causation ?
24. Feature Types
Feature
Variable
Categorical
Nominal lable ? Is Mother?
Ordinal Order
Which
child?
Numerical
discrete
continious
dosen’t
have
mathematic
meaning
Age No of
Child
Pain
Relief
Medic
ine
Food Temp Power Fuel
3 No No No 43C No No
25. Visualization Methods
• If you were to
• Tell your Story
• Analyze the Data
• Organize an Information
• OLAP Model
Distribution
Histogram
Sparse/Dense
Probable Range
Skew or Centered
Comparison
Bar Chart
(Multi) Category
vs Numerical
Box Plot
N-tile, Multi-Group
Composition
Part of Whole
Correlation
Scatter Plot
Making
sense
of
Data
scale
Normali
zing
Bucket
ing
27. Data Recognition
• Most Real-World Data are Wide , many Dimensions - Sparse
• Most Machine Data are Deep Few Dimensions
• Storage I/O .. less signatures
• Most Systems Data
• Chemical Factory.. Are more stable..
• Neural Data
• Image are Dense..
Wide
D
E
E
P
30. does data driven decision recommendation
make sense?
Exploratory
Mining
Insight
Does it expose a New reality ?
Does it Justify a Call for Action?
How strong is the reporting ?
How does it Vary from History to Present ?
Does it even require
31. XOP’s Bridging the Human <>AI
towards building a better future
with greater power comes greater responsibility.. let’s take some.. steps towards..
make it eco-friendly.. incorporate ethics.. construct explainable knobs.. Process for governance.. ensure
it’s secure.. make it approachable.. consumable. . understandable.. repeatable.. debuggable.. amenable..
whatever is your vision
Editor's Notes
https://www.quora.com/What-is-the-difference-between-feelings-and-emotions-1
Feelings, are generated in the heart and are related to one’s higher truth or Dharma. It is due to feelings that one does certain actions that uphold his true nature. Feelings are something you can choose from the inner depths of your soul. Feelings are the inner compass that help develop intuition.
Emotions are egoic.
News reporters : Live News of What's happening, Various aspects, Humanitarian, Pandemic, Travel, Medical ( Childbirth ) , Education
Historical Data : Taliban ( How it was handled and what it has impacted ?)
General Data : Refugee, War, Conflict
Question Needing Answers : Near Terms, ( Weeks ), Short Term ( Months), Mid Term( Quarter), Long Terms( Year)
UN : Release Fund or Not ?
Does it expose a New reality ?
Does it Justify a Call for Action?
How strong is the reporting ?
How does it Vary from History to Present ?
Does it even require