2008.05248.pdf (3.33 MB)
Null-sampling for interpretable and fair representations
conference contribution
posted on 2023-06-09, 21:28 authored by Thomas Kehrenberg, Myles Bartlett, Oliver Thomas, Novi QuadriantoNovi QuadriantoWe propose to learn invariant representations, in the data domain, to achieve interpretability in algorithmic fairness. Invariance implies a selectivity for high level, relevant correlations w.r.t. class label annotations, and a robustness to irrelevant correlations with protected characteristics such as race or gender. We introduce a non-trivial setup in which the training set exhibits a strong bias such that class label annotations are irrelevant and spurious correlations cannot be distinguished. To address this problem, we introduce an adversarially trained model with a null-sampling procedure to produce invariant representations in the data domain. To enable disentanglement, a partially-labelled representative set is used. By placing the representations into the data domain, the changes made by the model are easily examinable by human auditors. We show the effectiveness of our method on both image and tabular datasets: Coloured MNIST, the CelebA and the Adult dataset.
Funding
BayesianGDPR - Bayesian Models and Algorithms for Fairness and Transparency; G2903; EUROPEAN UNION
History
Publication status
- Published
File Version
- Accepted version
Journal
Computer Vision – ECCV 2020Publisher
Springer International PublishingPublisher URL
External DOI
Volume
12373Event name
European Conference on Computer Vision (ECCV)Event location
OnlineEvent type
conferenceEvent date
23 Aug 2020 - 28 Aug 2020ISBN
9783030586041Series
Image Processing, Computer Vision, Pattern Recognition, and GraphicsDepartment affiliated with
- Informatics Publications
Full text available
- Yes
Peer reviewed?
- Yes
Legacy Posted Date
2020-08-18First Open Access (FOA) Date
2021-12-20First Compliant Deposit (FCD) Date
2020-08-18Usage metrics
Categories
No categories selectedKeywords
Licence
Exports
RefWorks
BibTeX
Ref. manager
Endnote
DataCite
NLM
DC