Fair-doctor: Detecting and mitigating unfairness in neural networks

"Important decisions are increasingly based directly on predictions from classifiers; for example, machine learning models are now being used to facilitate autonomous vehicles, predict stock market trends, diagnose illnesses, and so much more. However, users fundamentally understand very little about how these black box classifiers come to make decisions, and whether these predictions are unbiased. With the more prevalent adoption of these systems, it is crucial that we must be able to both explain and understand what concepts our models utilize to make predictions to ensure that we are building unbiased, interpretable models. To facilitate this, we propose Fair-Doctor, a pipeline which diagnoses unfairness, treats it, and follows up to ensure that the algorithmic bias has been mitigated. We utilize TCAV (Testing With Concept Activation Vectors), a state-of-the-art interpretability tool, to diagnose unfairness. We also introduce a novel adversarial fairness loss function, which works to remove the specified unfairness in the model. We evaluate this architecture on a simple CNN trained on CelebA to predict how ""young"" a person looks. Our results demonstrate that we are able to successfully reduce the bias in this model."

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Adhikari, Rittika

Contributors dc:contributor

• Koyejo, Oluwasanmi

• Computer science

Statement dc:rights

• Copyright 2021 Rittika Adhikari

Language dc:language

en, eng

Handle dc:identifier

http://hdl.handle.net/2142/113921

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F113921