Is Fairness Only Metric Deep? Evaluating and Addressing Subgroup Gaps in Deep Metric Learning
Deep metric learning (DML) enables learning with less supervision through its emphasis on the similarity structure of representations. There has been much work on improving generalization of DML in settings like zero-shot retrieval, but little is known about its implications for fairness. In this pa...
Main Authors: | , , , , |
---|---|
Format: | Report |
Language: | unknown |
Published: |
arXiv
2022
|
Subjects: | |
Online Access: | https://dx.doi.org/10.48550/arxiv.2203.12748 https://arxiv.org/abs/2203.12748 |
id |
ftdatacite:10.48550/arxiv.2203.12748 |
---|---|
record_format |
openpolar |
spelling |
ftdatacite:10.48550/arxiv.2203.12748 2023-05-15T16:01:11+02:00 Is Fairness Only Metric Deep? Evaluating and Addressing Subgroup Gaps in Deep Metric Learning Dullerud, Natalie Roth, Karsten Hamidieh, Kimia Papernot, Nicolas Ghassemi, Marzyeh 2022 https://dx.doi.org/10.48550/arxiv.2203.12748 https://arxiv.org/abs/2203.12748 unknown arXiv arXiv.org perpetual, non-exclusive license http://arxiv.org/licenses/nonexclusive-distrib/1.0/ Machine Learning cs.LG Artificial Intelligence cs.AI Computers and Society cs.CY Machine Learning stat.ML FOS Computer and information sciences Preprint Article article CreativeWork 2022 ftdatacite https://doi.org/10.48550/arxiv.2203.12748 2022-04-01T18:13:00Z Deep metric learning (DML) enables learning with less supervision through its emphasis on the similarity structure of representations. There has been much work on improving generalization of DML in settings like zero-shot retrieval, but little is known about its implications for fairness. In this paper, we are the first to evaluate state-of-the-art DML methods trained on imbalanced data, and to show the negative impact these representations have on minority subgroup performance when used for downstream tasks. In this work, we first define fairness in DML through an analysis of three properties of the representation space -- inter-class alignment, intra-class alignment, and uniformity -- and propose finDML, the fairness in non-balanced DML benchmark to characterize representation fairness. Utilizing finDML, we find bias in DML representations to propagate to common downstream classification tasks. Surprisingly, this bias is propagated even when training data in the downstream task is re-balanced. To address this problem, we present Partial Attribute De-correlation (PARADE) to de-correlate feature representations from sensitive attributes and reduce performance gaps between subgroups in both embedding space and downstream metrics. : Published as a conference paper at ICLR 2022 Report DML DataCite Metadata Store (German National Library of Science and Technology) |
institution |
Open Polar |
collection |
DataCite Metadata Store (German National Library of Science and Technology) |
op_collection_id |
ftdatacite |
language |
unknown |
topic |
Machine Learning cs.LG Artificial Intelligence cs.AI Computers and Society cs.CY Machine Learning stat.ML FOS Computer and information sciences |
spellingShingle |
Machine Learning cs.LG Artificial Intelligence cs.AI Computers and Society cs.CY Machine Learning stat.ML FOS Computer and information sciences Dullerud, Natalie Roth, Karsten Hamidieh, Kimia Papernot, Nicolas Ghassemi, Marzyeh Is Fairness Only Metric Deep? Evaluating and Addressing Subgroup Gaps in Deep Metric Learning |
topic_facet |
Machine Learning cs.LG Artificial Intelligence cs.AI Computers and Society cs.CY Machine Learning stat.ML FOS Computer and information sciences |
description |
Deep metric learning (DML) enables learning with less supervision through its emphasis on the similarity structure of representations. There has been much work on improving generalization of DML in settings like zero-shot retrieval, but little is known about its implications for fairness. In this paper, we are the first to evaluate state-of-the-art DML methods trained on imbalanced data, and to show the negative impact these representations have on minority subgroup performance when used for downstream tasks. In this work, we first define fairness in DML through an analysis of three properties of the representation space -- inter-class alignment, intra-class alignment, and uniformity -- and propose finDML, the fairness in non-balanced DML benchmark to characterize representation fairness. Utilizing finDML, we find bias in DML representations to propagate to common downstream classification tasks. Surprisingly, this bias is propagated even when training data in the downstream task is re-balanced. To address this problem, we present Partial Attribute De-correlation (PARADE) to de-correlate feature representations from sensitive attributes and reduce performance gaps between subgroups in both embedding space and downstream metrics. : Published as a conference paper at ICLR 2022 |
format |
Report |
author |
Dullerud, Natalie Roth, Karsten Hamidieh, Kimia Papernot, Nicolas Ghassemi, Marzyeh |
author_facet |
Dullerud, Natalie Roth, Karsten Hamidieh, Kimia Papernot, Nicolas Ghassemi, Marzyeh |
author_sort |
Dullerud, Natalie |
title |
Is Fairness Only Metric Deep? Evaluating and Addressing Subgroup Gaps in Deep Metric Learning |
title_short |
Is Fairness Only Metric Deep? Evaluating and Addressing Subgroup Gaps in Deep Metric Learning |
title_full |
Is Fairness Only Metric Deep? Evaluating and Addressing Subgroup Gaps in Deep Metric Learning |
title_fullStr |
Is Fairness Only Metric Deep? Evaluating and Addressing Subgroup Gaps in Deep Metric Learning |
title_full_unstemmed |
Is Fairness Only Metric Deep? Evaluating and Addressing Subgroup Gaps in Deep Metric Learning |
title_sort |
is fairness only metric deep? evaluating and addressing subgroup gaps in deep metric learning |
publisher |
arXiv |
publishDate |
2022 |
url |
https://dx.doi.org/10.48550/arxiv.2203.12748 https://arxiv.org/abs/2203.12748 |
genre |
DML |
genre_facet |
DML |
op_rights |
arXiv.org perpetual, non-exclusive license http://arxiv.org/licenses/nonexclusive-distrib/1.0/ |
op_doi |
https://doi.org/10.48550/arxiv.2203.12748 |
_version_ |
1766397153985429504 |