2. Diversity in Data Science Space and Models
Vidhya Chandrasekaran
Women in Data Science | 9 March 2019 | IIM Bangalore
3. Some Stats to Start With
55%
Large Companies
that Adopted AI as
of Dec 2018
300,000
AI Engineers as
of Dec 2017
$60 billion
AI Revenue
by 2025
85%
Customer interactions
to be replaced by
AI in 2020
Gartner
5. Manual HR
Scroll through hundreds of profiles manually from different job sources
Large ETA, Human Bias, Error in Assessment
Screening and accessing the candidates skills against JD
Large ETA, Human Bias, Error in Assessment
Scheduling interview process with Recruiting manager
Human Bias, Scheduling Issues
Yahoo!!!
6. AI on HR
Rates and grades profiles that is more
relevant with the JD
Scheduling interview for top candidates
with the recruitment manager
• Always on HR
• Lesser turn around time for
recruitment
• Reduces Human bias in recruitment
• Better assessing the candidates
• Reduces Human bias in Promotions
• Fills the gap of technology skills in
HR
• Recommend the appropriate training
for the candidate
Yahoo!!!
7. Sexist Recruitment Tool?
• A ecommerce big player reportedly brought down its
internal AI recruitment tool for gender bias
• Gender-insensitive tool penalizes women candidates
since the data was trained from a workforce consisting
majorly of male employees
• Gave low grades to resumes with words like “Women”
• Candidates with a break in their career would be
considered a factor for low scoring, which impacts
women most
• “Executed”, “Captured” found in most male candidates
are scored high
• Instead of eliminating human bias, AI tool can bring in
unintentional bias
8. Challenges in Manual Law Enforcement
• Inadequate work force
• Human Bias
• High False Positives
• Human Bias
• Very Reactive approach
• High False Positives
• Workforce deployment
challenges
• Very Reactive approach
• Not enough technical
expertise to identify criminals
9. Law Enforcement with the Power of Data
• Proactive prevention of crime
• Better work force management
• Faster identification of criminals
• Optimizing security strategies
• Eliminate Intentional bias
• Identifying terrorism networks from
social media
10. AI Policing Gone Wrong
Stop and Frisk Program in New York uses AI
• Identifies 83% of Colored and Hispanic and 10% white people
• 52% black and Hispanic and 33% white population
ProPublica Report on racial bias in AI
Images courtesy: ProPublica
Subsequence offense
Dylan: 3 Drug possession
Bernard: None
Subsequence offense
Vernon: 1 Grand Theft
Brisha: None
11. Bias in Online Ads (Carnegie Mellon Research)
An experiment was conducted by Carnegie Mellon University with Adfisher.
• Exec jobs shown to 17% Women and 83% men
• 93% accurate that the model was biased
12. Skin color: Failure to recognize a human
Zip code: Denying same-day delivery
13. Mind Your Words
Men is to a Doctor as to Women is to a Nurse?
“Examining Gender and Racial Bias in Two Hundred Sentiment Analysis System” - Svetlana Kirichinko
and Saif Mohammad
• 75% to 86% of the sentiment analysis application scored sentences of one gender higher than the
other
• Associates Emotion with Women and Fear with Men
“Man is to Computer Programmer as Woman is to Homemaker” - Bolugbasi Tolga
• Found the following associations in the she-he analogies:
• Queen-king
• Sister-brother
• Daughter-son
• Mother-father
• Convent-monastery
• Waitress-waiter
• Nurse-physician
• Sewing-carpentry
• housewife-shopkeeper
• softball-baseball
• cosmetics-pharmaceuticals
• giggle-chuckle
• interior designer-architect
• charming-affable
14. The Infinite Loop – Accelerated Bias
Biased
Society
Data
from
Biased
Society
Model
Outputs
Action
Taken
Action
Recorded
15. Challenges in Eliminating Bias in AI
Vidhya Chandrasekaran
Women in Data Science | 9 March 2019 | IIM Bangalore
16. Fair AI
An AI can deliver a fair decision irrespective of
▪ Gender
▪ Age
▪ Race
▪ Educational background
▪ Economic background
▪ Physical appearance
▪ Power and authority …
What is fairness?
▪ Individual Fairness
o Similar people should be treated similarly
o Definition of similarity is the key
▪ Group Fairness
o Where fair decision making across groups split by
Race, Gender, Age, etc.
o Positive and Negative prediction rates equal across
groups
17. Why Bias
Accountability is difficult
No common law
Socio-Technical. Data
scientist think Loss function
Debugging and eliminating bias
is very complex
Difficult to eliminate hidden
bias from proxy variables that
has most information
Years of biased data
18. Can We Solve It?
Vidhya Chandrasekaran
Women in Data Science | 9 March 2019 | IIM Bangalore
19. Algorithmic Accountability
• Audits and standards should catch up with Technology
• Legal standards and practices should be established
• Having a governance on the algorithms
• Transparency and audit of code and logic
• Collaboration of Social scientist and Data scientist
• Compliances and governance on variable usage
• Platform to test fairness
20. Training Data
• Algorithms are fair to all only if they are trained on all
IBM research releases diversity in faces:
“For face recognition to perform as desired – to be both
accurate and fair – training data must provide sufficient balance
and coverage”
• Up-sampling and down-sampling and synthetic generation
• Transfer Learning where data inadequacy
• Universal unbiased word embeddings
Inclusive Training Sets
21. Decoupled Classifier – Group Fairness
“Decoupled Classifier for Group Fairness” -
Cynthia Dwork
Sensitive attributes classify with different
accuracy level for different classes
• Achieving both Fairness and accuracy is very difficult
• Build Multiple classifiers separate for each class
• One joint function to optimize
• Output is chosen based on minimizing the joint
function
• False positives are same across groups
23. Regularization to Treat Bias
Penalize if variables are given
more importance
Dropout nodes randomly to
avoid reliance
24. Randomness to the Rescue
“Accountability of any process should grapple it” – Joshua A. Kroll in
Accountable Algorithms (On randomness to design of a process)
• Many Machine learning algorithm incorporates Randomness
• Randomize the data used for training
• Randomize the features used for the model to avoid weighing heavily on one
particular feature
• Assign a different class value to the positive and negative classes predicted
for random sample