This document discusses bias in machine learning algorithms and datasets. It notes that high bias in algorithms can cause them to miss important relationships, and that datasets are often not standardized or representative. Examples given include facial recognition algorithms performing worse on dark-skinned women, and ads displaying higher-interest credit cards to black users at a higher rate. The document calls for assessing whether problems need machine learning solutions, testing models on diverse data, being open to criticism of models, and assuming bias will persist until steps are taken to address it.
2. SOCIETY
Prejudice in favor of or against one thing,
person, or group compared with another,
usually in a way considered to be unfair.
The bias error is an error from erroneous
assumptions in the learning algorithm.
High bias can cause an algorithm to miss the
relevant relations between features and
target outputs (underfitting).
MACHINE LEARNING
7. IGNORANCE IS BIAS
It shouldn't need be explained how it's
ethically wrong to attempt to predict a
person's sexuality based on appearance.
8. FAILING PEOPLE OF COLOR
Many widely used facial reconigtion
algorithms by Microsoft Amazon, and
Face++ performed 35% worse on dark
skinned women.
In 2017 mulitple Chinese users reported
being able to log into another's iPhone.
using FaceID.
9. INNACURACIES AND
INCONVENIENCE
There is bias in travel for people who may be
considered immigrants to some, that will be
exemplified with automated boarding.
12. Algorithms with little feedback have
little opportunity to "learn"
Algorithms can reflect xenophobic bias
and use limited examples to misidentify
people with dark skin as criminal.
With few examples of successful minority
candidates predection based on these
imbalanced sets will amplify bias.
Online ads that determine a user's race
to be black display ads for high interest
credit cards at a higher rate than others.
14. BIASED DATA = BIASED OUTCOMES
Data used for training models is hardly ever
actually representative of people it will be used
on.
HUMANS ARE IMPERFECT
This doesn't mean algorithms unbiased, it means
we have to assume bias will persist until we take
steps to remove it.
MASTERS OF OUR DEMISE
We are in one of few fields where we get to pick
the metric we're measured against.
15. COMPAS
Consistently ranks black and brown
incarcerated people as more likely to offend
and higher risk than white prisoners.
FAILED SENSORS
Milimeter wave sensors used by the TSA
consistently have trouble with black
women's hair causing travel delays.
FLAWED HARDWARE
Camera hardware has been tuned and
developed to highlight lighter skin tones.
18. ASSESING THE NEED
Not all problems are best solved with
machine learning.
TEST WITH EDGE CASES
Hardware should be tested on dark
skinned first.
EXTREME SKEPTICISM
Assesing models with harsh criticism
towards performance on edge cases.
19.
20. HOW TO START A ML PROJECT
(ETHICALLY)
QUESTION IF THE
SOLUTION FITS
THE PROBLEM
EXAMINE FOR AND
REMOVE BIASED
PROXIES
GIVE YOUR MODEL
FEEDBACK AND
TEST FOR BIAS
25. ZIP CODE
Applying for a credit card with a 90210 zip code
shouldn't improve your chances of getting
approved.
HOMOGEN OUS EXAMPLES
If there is an extreme imbalance between
classes any model can attribute an occurence to
subtle proxies the model learns.
WORD EMBEDDINGS
Tools used to measure the distance of the
meaning of words can embed our sexist cultural
norms and exclude nonbinary people.
26. MINDSET CHANGE
Assume there will be some aspect of bias in your
models and asses potential consequences.
ACCOUNTABIL IT Y
Build user trust and be open to algorithmic
criticism by making models open source.
Transparency leads to accountability,
WORD EMBEDDINGS
Tools used to measure the distance of the
meaning of words can embed our sexist cultural
norms and exclude nonbinary people.