federica bianco
astro | data science | data for good
University of Delaware
Department of Physics and Astronomy
federica bianco
she/her
Data Science Institute
Biden School of PublicPolicy & Administration
UD Center for Science, Ethics, and Public Policy
Ethics and diversity in AI and ML
This deck
fbianco@udel.edu
NASA has the opportunity to set up its AI operations embedded in an ethical framework
Organizations inspect their ethical responsibility after incidents
from
Ethics of Science
to
Ethics of AI
The Guidelines on Ethics from the American Physical Society state: As citizens of the global community of science, physicists share responsibility for its welfare.
The Dual-use-dilemma: Research delivers knowledge that can be used for unintended purposes, including illegal and societal- and ethically-disputable applications. The responsibility of the tools’ creator has long been the subject of debate. The same is true for innovation in data science, machine learning, and artificial intelligence.
Vinay Prabhu
Who is responsible?
Tool creator?
AI developer
Practitioner that selects the tool?
Sheriff/DA office?
Decision maker that interprets the tools results?
Police?
Ethics of AI
moral implications of the technology we label AI.
Information philosophy
Data Ethics
Futurism
Scify
machine ethics
is concerned with ensuring that the behavior of machines toward human users
Human-Robot Interaction
is concerned with ensuring that the behavior of machines toward human users
Privacy and Surveillance
collection access to information
Algorithmic bias and trust in AI
embedding bias and opacity of ML algorithm
Ethics of AI
Information philosophy
Data Ethics
Futurism
Scify
machine ethics
is concerned with ensuring that the behavior of machines toward human users
isaac asimov laws of robotics, 1942
Ethics of AI
Information philosophy
Data Ethics
Futurism
Scify
Human-Robot Interaction
Blade Runner, Ridley Scott, 1982
Sophia Robot, Citizen of Saudi Arabia, 2016
David Hanson |
Ethics of AI
Privacy and Surveillance
Information philosophy
Data Ethics
Futurism
Scify
Who collects the data, Who accesses the data?
Balance between Open data, reproducibility, and preserving privacy?
NASA RELEVANCE: imaging satellites, GPS
Ethics of AI
Diebold, Inc., has crossed over the digital threshold with new observation technology that the company believes will transform the surveillance world ... Diebold's answer to analog stemmed from a Space Act Agreement with NASA's Glenn Research Center, in which the North Canton, Ohio-based company acquired the exclusive rights to video observation technology that was designed for high-speed applications and does not require human intervention.
Ethics of AI
Algotirhmic Bias and Trust in AI
Information philosophy
Data Ethics
Futurism
Scify
ML models "learn" from data examples by minimizing some an objective function.
ML models "learn" from data examples by minimizing some an objective function.
Information philosophy
Data Ethics
Futurism
Scify
It can learn and amplify bias in data
ML models "learn" from data examples by minimizing some an objective function.
(covarience is gonna get you)
If the data is biased, model learn and amplify the bias
1.
Ethics of AI
Algorothmic Bias
ML is a representation of the "world"
what world?
the world we live in?
or the world we would like to live in?
who decides what that looks like?
Ethics of AI
Trust in AI and AI opacity
Information philosophy
Data Ethics
Futurism
Scify
DARPA’s Explainable Artificial Intelligence Program Gunning & Aha
generalized additive models
decision trees
SVM
Random Forest
Accuracy
univaraite
linear
regression
Deep NeuralNets
univaraite
linear
regression
If model cannot be interpreted can they be trusted?
can they fulfill the Right to Explanation?
interpretability
Ethics of AI
Trust in AI and AI opacity
Information philosophy
Data Ethics
Futurism
Scify
DARPA’s Explainable Artificial Intelligence Program Gunning & Aha
trivially intuitive
generalized additive models
decision trees
SVM
Random Forest
Accuracy
univaraite
linear
regression
we're still trying to figure it out
Deep NeuralNets
univaraite
linear
regression
interpretability / time
If model cannot be interpreted can they be trusted?
can they fulfill the Right to Explanation?
Information philosophy
Data Ethics
Futurism
Scify
Deep NeuralNets
generalized additive models
decision trees
Random Forest
Accuracy
univaraite
linear
regression
interpretability / time
univaraite
linear
regression
univaraite
linear
regression
univaraite
linear
regression
SVM
Ethics of AI
Trust in AI and AI opacity
number of features
If model cannot be interpreted can they be trusted?
can they fulfill the Right to Explanation?
Information philosophy
Data Ethics
Futurism
Scify
Deep NeuralNets
generalized additive models
decision trees
Random Forest
Accuracy
univaraite
linear
regression
interpretability / time
univaraite
linear
regression
univaraite
linear
regression
univaraite
linear
regression
SVM
Model selection need to balance interpretability with accuracy
2.
Ethics of AI
Trust in AI and AI opacity
number of features
Ethics of AI
Fairness in AI
Information philosophy
Data Ethics
Futurism
Scify
What should be maximized?
ML models "learn" from data examples by minimizing some an objective function.
the hypothetical trolley problem suddenly is real
Ethics of AI
Fairness in AI
Information philosophy
Data Ethics
Futurism
Scify
Public safety vs unfair punishment
A fully harmless objective function may not exist. How do we balance potential harm?
3.
https://dl.acm.org/doi/abs/10.1145/3442188.3445922
On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜
Bender, Gebru et al 2021 “It is past time for researchers to prioritize energy efficiency and cost to reduce negative environmental impact and inequitable access to resources”
A framework for enabling AI within ethical boundaries
The study of ethical implications of the creation, use, and coexistence with autonomous systems
Responsible AI is the practice of designing, developing, and deploying AI with good intention to empower employees and businesses, and fairly impact customers and society
AI testing
Adversarial Robustness and Privacy
Fairness, Accountability, Transparency
Fairness, Accountability, Transparency
Uncertainty quantification
Explainable AI
The Next Generation Internet initiative by the Digital Single Market of the European Commission.
https://ngi.eu/wp-content/uploads/sites/48/2018/07/Responsible-AI-Consultation-Public-Recommendations-V1.0.pdf
Starting now, NASA has the opportunity to set up its AI operations embedded in an ethical framework
Set up AI operations embedded in an ethical framework
ASK:
What is the decision process?
Who needs to be involved?
What are the values that we should be encoding in the processes that leads to build, deploy, and use of AI?
"Human-Centered AI"
What is the decision process?
Ben A. Schneiderman - UMD
"Human-Centered AI"
What is the decision process?
I can build it, but should I build it?
Who will be impacted?
Who will be accountable?
Whose job will it replace?
Whose jobs will this enable?
Who is the new workforce and how are they trained?
yes, but WHO is testing?
Who needs to be involved?
Diversity is key to identify potential harm
"Human-Centered AI"
Why does this AI model whitens President Obama's face?
Simple answer: the data is biased. The algorithm is fed more images of white people
Complex answer: the testing and validation framework validated the bias
model: PULSE (GAN)
Who needs to be involved?
(suggested by me.... please do consult many other experts and don't take it as an exhaustive list!!)
papers and reports
The National Security Commission on Artificial Intelligence
The National Security Commission on Artificial Intelligence
https://www.nscai.gov/
Ben Schneiderman 2020 "Human-Centered Artificial Intelligence: Trusted, Reliable & Safe" http://www.cs.umd.edu/hcil/trs/2020-01/2020-01.pdf
Bender et al. 2021 On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜https://dl.acm.org/doi/abs/10.1145/3442188.3445922
Stanford Encyclopedia of Philosophy https://plato.stanford.edu/entries/ethics-ai/#PrivSurv
https://ai.google/
programs, plans, and others
Dara Responsibly Comics https://dataresponsibly.github.io/comics/
Microsoft Responsible AI plan https://docs.microsoft.com/en-us/learn/modules/responsible-ai-principles/4-guiding-principles
Google Responsible AI plan https://ai.google/responsibilities/responsible-ai-practices/
DARPA’s Explainable Artificial Intelligence (XAI) Program https://ojs.aaai.org/index.php/aimagazine/article/view/2850
An interview with PwC Maria Luciana Axente https://www.datacamp.com/community/blog/the-future-of-responsible-ai
PULSE: https://github.com/adamian98/pulse#what-does-it-do
GTP-3 https://arxiv.org/abs/2005.14165v4
GTP-3 text generator https://deepai.org/machine-learning-model/text-generator
COMPAS (wiki entry, no paper) https://en.wikipedia.org/wiki/COMPAS_(software)
models mentioned in the talk
By federica bianco
NASA ML/AI retreat, presentation on Ethics and diversity in ML and AL - Jan 2022