Computer Vision MT25, Ethics and privacy
Flashcards
Why is it not sufficient to omit sensitive features (e.g. race) from an ML model used to make decisions?
Other features may correlate with the sensitive features.
Suppose:
- $Y$ is a target variable (e.g. recidivism)
- $R$ is the output of a classifier
- $A$ is a sensitive attribute
@Define the fairness-related “independence” condition in this context.
The classifier response is independent from the sensitive attribute, i.e.
\[\mathbb P(R \mid A) = \mathbb P(R)\]Suppose:
- $Y$ is a target variable (e.g. recidivism)
- $R$ is the output of a classifier
- $A$ is a sensitive attribute
@Define the fairness-related “separation” condition in this context.
The classifier response is conditionally independent from the sensitive attribute given the target, i.e.
\[\mathbb P(R, A \mid Y) = \mathbb P(R \mid Y) \mathbb P(A \mid Y)\]Suppose:
- $Y$ is a target variable (e.g. recidivism)
- $R$ is the output of a classifier
- $A$ is a sensitive attribute
Then the fairness-related “separation” condition in this context is that the classifier response is conditionally independent from the sensitive attribute given the target, i.e.
\[\mathbb P(R, A \mid Y) = \mathbb P(R \mid Y) \mathbb P(A \mid Y)\]
What does this entail in terms of the error rates of the classifier?
All groups (delineated by $A$) experience the same false negative and false positive rate.
Suppose:
- $Y$ is a target variable (e.g. recidivism)
- $R$ is the output of a classifier
- $A$ is a sensitive attribute
Then we have the fairness conditions:
-
Independence: $\mathbb P(R, A) = \mathbb P(R) \mathbb P(A)$
-
Separation: $\mathbb P(R, A \mid Y) = \mathbb P(R \mid Y) \mathbb P(A \mid Y)$
@State an unfortunate theorem in this context.
Suppose further:
- $Y$ is binary
- $A$ is not independent of $Y$
- $R$ is not independent of $Y$
Then:
- Both independence and separation cannot hold simultaneously.
@Define allocative harms in the context of ML-decision making.
Harms causes by a system allocating resources unfairly.
@Define representational harms in the context of ML-decision making, and give 5 distinct types of such harms.
Where a system reinforces harmful stereotypes.
- Denigration: Use of culturally disparaging terms.
- Stereotype: Reinforces negative stereotypes.
- Recognition: A group is erased or made invisible.
- Under-representation: A group is under-represented.
- Ex-nomination: Representing ideology as common sense.
Can you classify these representational harms into the (potentially simultaneous) categories of:
- denigration
- stereotype
- recognition
- under-representation
- ex-nomination
