AI, machine learning, robotic processing, and automation introduce risk and bias that may have profound and specific impacts on customers and users. We need to invest in data and tools to facilitate the ethical use and management of automated prediction applications. We need to fill data gaps and build AI Ops, Privacy, Security, and Life Cycle Management.
4. PickAxes & Shovels LLC Confidential
“We are flying blind”
in our conversations and decision-making about AI
Stanford University 100 Year Study on AI
6/2/19 4
75% of CEOs
say potential for bias and
lack of transparency
impede AI adoption
- PwC CEO Pulse Survey 2017
Only 38% of
Firms
have a clearly designed
strategy for
implementing AI
Boston Consulting Group 2019
5. Spending on Artificial Intelligence (AI) is accelerating
5
$5.3B
$7.8B
$12.38B
$19.1B
$29.8B
$46.B
$52.2B
$0
$15
$30
$45
$60
2015 2016 2017 2018 2019E 2020E 2021E
Worldwide Spending on cognitive and artificial intelligence systems
Source: IDS Research 2017
PickAxes & Shovels LLC Confidential6/2/19 5
7. Regulatory push is heating up
• Requires intelligible consent, access & portability, right to be forgotten, privacy by design,
and appointment of data protection officer
• Fine is Euro20M or 4% of annual turnover
• GDPR enforcement of data breaches is growing since it went into effect in May 2018
• 144,376 complaints submitted by people who feel their privacy has been impacted
• 89,271 data breaches (72 hours to report)
• Multiple enforcement actions , most small– among the largest: $50M (Google),
$400K, $220K, $20K
European Union’s GDPR
PickAxes & Shovels LLC Confidential6/2/19 7
8. Regulatory push is heating up
US Algorithmic Accountability Act in the House of Representatives
PickAxes & Shovels LLC Confidential6/2/19 8
9. Consumer trust is at risk - not only data security - but
also provenance and permissions
PickAxes & Shovels LLC Confidential6/2/19 9
10. Bias in Bias out (BIBO) creates legal and brand risk
Discrimination Allocation
11. You don’t want your company to have a
Face++, Amazon, or Nikon Moment
Face++, Amazon Nikon
PickAxes & Shovels LLC Confidential6/2/19 11
12. Automation amplifies
400,000 views in days for
video of girls playing at pool
YouTube recommendation
algorithm seeds content
to pedophiles by noting
affinity for young partially clad
children
PickAxes & Shovels LLC Confidential6/3/19 12
YouTube’s Recommendation AI
Drives 70% of views
The company does not reveal
how the system makes its
choices
13. Model and Hardware management add complexity
PickAxes & Shovels LLC Confidential6/2/19 13
16. Putting ethics into practice
Privacy and Trust
• Informed Consent, Control, and Choice
• Responsible Use & Data Minimization
Transparency
• Interpretable Models & Techniques
• Explainable Layers
16PickAxes & Shovels LLC Confidential6/2/19
Fairness
• Representative data and data scientists
• Human centered, accessible design
Accountability
• Auditability
• Responsiveness
17. These seem more like guidelines, how
do we make this actionable?
18. Make investments to support ethical prediction
Data:
identify, aggregate,
scale data sets to fill
gaps
AI Ops:
Provenance,
privacy,
transparency,
security,
management
PickAxes & Shovels LLC Confidential6/2/19 18
19. Bias in bias out (BIBO) risk
Model A
Predictions with
unintended
consequences
Data Gap
Data Gap
PickAxes & Shovels LLC Confidential6/2/19 19
21. Changes in data or model updates alter prediction results
Model A PredictionsData A
Model A1
Different
PredictionsData A +
New Data
PickAxes & Shovels LLC Confidential6/2/19 21
22. Changes in hardware alter prediction results
Model A Hardware A
Accurate
Predictions
Model A Upgrade or
Hardware B
False Positives
or False Negatives
PickAxes & Shovels LLC Confidential6/2/19 22
23. Invest in tools to manage data and
predictive applications over their
lifecycle
24. Some questions to ask:
• How do we intend to use the data we collect? Can we explain this to consumers in plain language?
• Do we practice data minimization? Or, are we blanket collecting? If so why? Balance business
opportunity of future uses with risk of holding the data
• Where did the data come from? How are we tracking provenance and permissions for use when opt-ins
change frequently
• Who is on our data science team? Who is building models?
• Do we have a process for evaluating data input, model choices, hardware, administration, software and
hardware updates and versioning? Do we have people, infrastructure, tools?
• Are we using interpretable models where transparency is required? If not, are we using techniques or
tools for interpretability and/or explainability? Are these auditable?
• Could we answer a customer’s question about how or why a decision was reached?
• How are we evaluating bias, fairness, and inclusion?
• How will we manage mistakes?
PickAxes & Shovels LLC Confidential6/2/19 24
25. Company Sector Capital Raised Investors
Genomics
$112M
$80M Series C 09/17
$45M Series B 09/16
General Catalyst, CRV,
Emerson Collective,
Khosla Ventures
Fintech
$18.2M
$12.8M Series A 12/18
NYCA Partners, Omidyar
Group, Kleiner Perkins
Explainability
$12M
$8.5M Series B 08/17
Pivot North Capital, Darling
Ventures, Citrix Systems
Enterprise AI Management
$37.9M
$25M Series B 05/19
$10.5M Series A 06./17
Norwest Venture Partners,
Gradient Ventures,
Madrona Venture Partners
VC beginning to recognize opportunity
25PickAxes & Shovels LLC Confidential6/2/19 25
26. A few resources
Organizations & Companies
• Omidyar: Race to the Top & Ethical OS
• Open Data Institute: Data Ethics Canvas
• AI Now Institute
• IEEE
• Algorithmic Justice League
• FATE ML: Fairness Accuracy Trust Ethics
• IBM 360 Fairness Tools
• Google TCAV and What-if-Tool
• Facebook Fairness
• Algorithmia
Open Source Tools & Techniques
• AI Open Scale
• Lime & Deep LIFT
• Layerwise Relevance Propogaton (LRP)
• GANs
• RETAIN – Georgia Tech method to help medical
professionals understand model predicting heart
failure
• RISE – Boston University model makes heat maps
to show what parts of neural network active in
prediction
• Active Fairness – MIT Media Lab Noriega & Bakker
paper
PickAxes & Shovels LLC Confidential6/2/19 26