Lab presentation (a framework for understanding unintended consequences of machine learning)
1. A Framework for Understanding
Unintended consequences of
Machine Learning
Author: Harini Suresh (MIT), John V. Guttag(MIT)
Presented: Chenguang Xu “Shine”
2. The Problem with Biased data
• Various unwanted consequences of ML algorithm arise in
some way from biased data.
• Bias refers to an unintended or potentially harmful
property of the data.
• Data is a product of many factors, and is the product of a
process
3. An illustrative Scenario
Lack of data on women, introducing
more data solved the issue.
The use of a proxy label (human assessment of
quality) versus the true label (actual qualification)
allowed the model to discriminate by gender.
5. Historical Bias
It is a fundamental, structural issue with the
very first step of the data generation process.
6. Representation Bias
• It arises when defining and sampling from a population.
• It can arise for several reasons:
• The sampling methods only reach a portion of the
population.
• The population of interest has changed or is distinct
from the population used during model training.
7. Representation Bias (cont.)
Shankar, Shreya, et al. "No classification without representation: Assessing geodiversity issues in open
data sets for the developing world." arXiv preprint arXiv:1711.08536 (2017).
8. Representation Bias (cont.)
Photos of bridegrooms from
different countries aligned by the
log-likelihood that the classifier
trained on Open Images assigns to
the bridegroom class.
Shankar, Shreya, et al. "No classification without representation: Assessing geodiversity issues in open
data sets for the developing world." arXiv preprint arXiv:1711.08536 (2017).
9. Measurement Bias
• It arises when subsequently choosing and measuring the
particular features of interest.
• It can arise in several ways:
• The granularity of data varies across groups.
• The quality of data varies across groups.
• The defined classification task is an oversimplification.
10. • It arises when a one-size-fit-all model is used for groups
with different conditional distributions.
Aggregation Bias
11. Evaluation Bias
• It occurs when the evaluation and/or benchmark data for
an algorithm doesn’t represent the target population.
Buolamwini, Joy, and Timnit Gebru. "Gender shades: Intersectional accuracy disparities in
commercial gender classification." Conference on Fairness, Accountability and Transparency.
2018.
12. Formalizations and Mitigations
• A data generation and ML pipeline viewed as a series of
mapping functions.
Mitigating Aggregation Bias:
• adjusting g
• change r or t for transforming
the data
Mitigating Evaluation Bias:
• redefine k
• adjusting X, Y
^ ^
Mitigating Representation Bias:
• improve s
Measurement and historical Bias:
• adjust s will likely be ineffective