Responsible AI

AI & LLMsMachine Learning

Materials — open to everyone, no sign-in

Topic: Responsible AI

Presenter: Ying Ying Liu

Additional Resources:


System Design Presentation - Responsible AI

Join Us on Wechat

Subscribe to Our YouTube channel https://commitway.com/eventyoutube

2 key points in responsible AI

Privacy

Fairness

2 definitions of responsible AI

China’s law and regulation on AI

United States’ law and regulation: HIPPA etc

Canada: more detailed regulation than United States, at national, provincial level

China is most advanced in law development

US&Canada: more regulation for government use of AI

Cases

Royal free, deepmind collaboration

Deepmind received personally identifiable information

Clearview AI

Collection of images online

Used by law enforcement

2020 Canada investigated Clearview AI collection methods

Deep fake

Emma

Taylor Swift

Canadian regulation and advices

Legal authority and consent

Appropriate purposes

Necessity and proportionality

Openness

Accountability

Individual access - user can see the process

Limiting collection, use and disclosure

Accuracy

Safeguard

Microsoft AI privacy

Differential privacy

grouping/aggregation

Anonymity vs utility

Adding Laplace distribution noise to raw data

Can be controlled by epsilon

Epsilon increases, utility increases, privacy decreases

Used by Apple, Google and Microsoft

The model performs well.

Need intuition

Need to collaborate across teams and companies

Understand laws and regulations

Feature: name, age, address, occupation, etc

Understand the source of the data

Try to have minimum usage

Should document different points that touch the data

Need to adjust epsilon

Collaboration

Transparent note (microsoft), privacy assessment (canadian government)

Fairness

Unfair treatment

Transparency

Cases for fairness

2007 Amazon AI hiring tool discriminated

Gender classification of people of darker skin color

2019 Racial bias in health scores

Labeling, using healthcare cost as measurement of health

Enhancement: use a combination of health and health care cost

[why not remove the healthcare cost?]

[ Probably due to the fact it’s for health insurance ]

2023 AI index report

LLM shows gender bias

2023 Quebac sues Facebook for ads to discriminate against age and gender

FBPML (foundation for best practices in machine learning)

Fair AI

Currently AI fairness is only on the surface

Religion

Fair machine learning

De-bais data

Fairlearn

How to improve fairness

Genetic programming

Generic algorithm - fixed array

Genetic programming - evolve the code. Using tree

Find optimal algorithm

Symbolic regression

[huge search space?]

Fair ML through Genetic coding

Objective 1: accuracy

Objective 2: fairness

Summary

Fairness is not mature yet, still a huge challenge

Privacy is more mature

Countries may learn from China

Developers:

Should be sympathetic

Follow best practice and tools

Understand the transparency

Transparency note - product deliverable of

[what are typical deliverable]

Human in the loop - subject matter expert can review the process

Structural injustice - human society is unfair - cannot be completely fair, cannot be 100%

Architecture diagram

Transparency note: a master document - small documents (privacy impact assessment)

Audience

Q: Is unfairness due to structural injustice or error in data?

A: There are some standards to measure. Gender and ethnicity

Q: What other aspects are mature: Transparency, useability

A: reliability, safety, security are more mature. Technically solvable problem is more mature

Q: genetic programming

A: Tree node: operator. Leaf node: are input data

A tool for tuning. May accomplish simple tasks

Q: transparency note, anything else?

A: Algorithm design

A collection of documents.

Human in the loop. Who to collaboration

Why fairness? Add business rule Customer service

Privacy: privacy impact assessment. Long document at provincial level

Security

Accountability: racy - matrix who is responsible for what.

Transparency note - workflow for the algorithm

Master document

Operation guide

End-user guide

Budget / expenditure

Most important is security and privacy question

Lots of paperwork for data science projects

Q: do different departments have different opinion about fairness

A: yes. Need communication across departments and compromises