Lecture 25: Ethics in Machine Learning

“With great power comes greater responsibility”.

Learning goals: - How to measure fairness - How to mitigate biases

1. Framework of harms

We analyze different types of harms caused by algorithms:

Quality-of-service harms
Distributive harms
Existential harms

2. Detecting hidden biases

Algos are hypersensitive to data that they are trained on. They might recognize patterns that we cannot recognize easily. See slide 40.
- Make this a paragraph in PWR paper.
  - Also include model cards initiative (i.e. in transparency section).
  - The section on Part 1: Framework of Harms can be used to show the effects biases have on society at greater.
Case study: st. georges algorithm: i.e. NLP for ethnic names

3. Algorithmic fairness

Adapting philosophical terms to ai algorithms.