1 of 18

What is AI Bias

Susama Agarwala

JHU APL

2 of 18

Why a seminar on data science?

3 of 18

What is data science?

4 of 18

What is AI?

In this course:

Algorithms that makes predictions/classifications based off a given data set.

Many algorithms not considered:

E.g.

Reinforcement learning algorithms

Human machine teaming algorithms

5 of 18

What is AI good for?

Protein Folding

Chess, Go and other games

Atari, Starcraft and other video games

6 of 18

Use cases

  • Credit scoring
  • Hiring decisions
  • Assistive medical technology
  • Facial recognition
  • Object recognition
  • Ad/outreach targeting
  • Protein folding
  • Climate change modeling
  • Voice recognition
  • Performance analytics
  • Machine breakdown prediction
  • ……..

7 of 18

What is it bad for?

8 of 18

What is going on?

Simple example: Linear Regresssions

9 of 18

What is going on?

Good application

Bad application

  • Credit scoring
  • Hiring decisions
  • Self driving cars
  • Assistive technology in medical technology
  • Facial recognition
  • Object recognition
  • Route planning
  • Ad/outreach targeting
  • Advanced Logistics
  • Voice recognition
  • Performance analytics
  • Machine breakdown prediction
  • ……..

Protein Folding

Atari, Starcraft and other video games

Chess, Go and other games

10 of 18

What happens if the training data is not i.i.d. distributed from the real world data?

11 of 18

What happens if key groups are under represented in the data?

12 of 18

What happens if there are differential error rates?

13 of 18

What happens if one fails to identify all the relevant subgroups?

14 of 18

What happens if a subpopulation has different characteristics?

15 of 18

Demo: CelebA dataset

15

24 January 2022

16 of 18

What makes a celebrity attractive?

16

24 January 2022

17 of 18

What ELSE makes a celebrity attractive?

17

24 January 2022

18 of 18

Is it really picking up annotator’s racial bias?

18

24 January 2022

This is a good example of what can happen when using this dataset to train a classifier:

https://medium.com/@mike.leske/how-i-accidentally-created-a-racist-ai-by-a-naive-dataset-selection-2cc9bf369bfa