pith. sign in

arxiv: 2603.05582 · v2 · pith:CGZ3UKUOnew · submitted 2026-03-05 · 💻 cs.LG · cs.CV

Bias In, Bias Out? Finding Unbiased Subnetworks in Vanilla Models

Ivan Luiz De Moura Matos , Abdel Djalil Sad Saoud , Ekaterina Iakovleva , Vito Paolo Pastore , Enzo Tartaglione This is my paper

Pith reviewed 2026-05-15 16:09 UTC · model grok-4.3

classification 💻 cs.LG cs.CV
keywords bias mitigationsubnetwork extractionpruningfairnessdebiasingdeep learningparameter removal
0
0 comments X

The pith

Standard neural networks trained on biased data already contain unbiased subnetworks that can be isolated by pruning without retraining.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that conventionally trained deep learning models already hold subnetworks that avoid biased features. A pruning method called BISE identifies and extracts these subnetworks directly from the original parameters. This extraction requires no extra unbiased data and no finetuning, yet the resulting subnetworks keep task performance while reducing reliance on biased cues. If correct, bias mitigation becomes a matter of parameter removal rather than full retraining or dataset redesign.

Core claim

Bias-Invariant Subnetwork Extraction (BISE) locates and isolates bias-free subnetworks that already exist inside conventionally trained models. These subnetworks are obtained through pruning and run without any parameter changes, relying less on biased features while preserving robust accuracy on standard benchmarks.

What carries the argument

Bias-Invariant Subnetwork Extraction (BISE): a pruning procedure that selects and retains only the parameters forming bias-free subnetworks within a vanilla-trained model.

If this is right

  • Extracted subnetworks rely less on biased features while keeping task performance.
  • Bias mitigation occurs through parameter removal rather than retraining or data changes.
  • The approach works on pre-trained models without additional unbiased training sets.
  • Resulting models are more computationally efficient than methods that retrain all parameters.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Bias encoding may be localized to specific parameter subsets rather than distributed uniformly.
  • The same pruning idea could be tested on other model properties such as robustness to distribution shift.
  • Post-training fairness adjustments become feasible for already deployed networks.
  • Training dynamics might be re-examined to see how biases concentrate during optimization.

Load-bearing premise

Bias-free subnetworks already exist inside models trained on biased data and can be reliably found by pruning without any unbiased examples or retraining.

What would settle it

Finding that every pruned subnetwork extracted by BISE still shows the same bias levels as the full original model on the tested benchmarks would falsify the claim.

Figures

Figures reproduced from arXiv: 2603.05582 by Abdel Djalil Sad Saoud, Ekaterina Iakovleva, Enzo Tartaglione, Ivan Luiz De Moura Matos, Vito Paolo Pastore.

Figure 1
Figure 1. Figure 1: Overview of the proposed method. BISE aims to extract [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of BISE. Solid black arrows indicate for [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Analysis of (a) coefficient γ and (b) pruning strategies. Shaded areas indicate the interval of one standard deviation. weights [29]), random pruning, and BISE (here, filters are ranked and removed according to their corresponding mi). We show results for different sparsity levels. It is important to highlight that we do not finetune the models considered. For random pruning, accuracy rapidly drops as the … view at source ↗
read the original abstract

The issue of algorithmic biases in deep learning has led to the development of various debiasing techniques, many of which perform complex training procedures or dataset manipulation. However, an intriguing question arises: is it possible to extract fair and bias-agnostic subnetworks from standard vanilla-trained models without relying on additional data, such as unbiased training set? In this work, we introduce Bias-Invariant Subnetwork Extraction (BISE), a learning strategy that identifies and isolates "bias-free" subnetworks that already exist within conventionally trained models, without retraining or finetuning the original parameters. Our approach demonstrates that such subnetworks can be extracted via pruning and can operate without modification, effectively relying less on biased features and maintaining robust performance. Our findings contribute towards efficient bias mitigation through structural adaptation of pre-trained neural networks via parameter removal, as opposed to costly strategies that are either data-centric or involve (re)training all model parameters. Extensive experiments on common benchmarks show the advantages of our approach in terms of the performance and computational efficiency of the resulting debiased model.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces Bias-Invariant Subnetwork Extraction (BISE), a pruning-based method to identify and isolate bias-free subnetworks that purportedly already exist inside conventionally trained (vanilla) models. These subnetworks are claimed to be extractable without any additional unbiased data, without retraining or fine-tuning, and to maintain competitive performance while relying less on biased features. Experiments on standard benchmarks are said to demonstrate advantages in both accuracy and computational efficiency relative to data-centric or full-retraining debiasing approaches.

Significance. If the central claim holds—that bias-free subnetworks can be reliably isolated from vanilla models using only the original biased training data—the result would be practically significant. It would offer a low-cost structural debiasing route that avoids both dataset curation and parameter updates, which is attractive for large pre-trained networks. The work also supplies a concrete test of the “lottery ticket” style hypothesis in the fairness setting.

major comments (2)
  1. [§3] §3 (BISE algorithm): the mask-selection objective is not shown to be free of an external bias signal. The description states that pruning uses only the original training loss, yet the selection must still quantify “reliance on biased features.” Without an explicit equation or pseudocode step that defines this quantification (e.g., a gradient-based attribution term or a fairness regularizer), it is impossible to verify that the procedure avoids the very information it claims to forgo.
  2. [§4.2] §4.2 (experimental protocol): the reported fairness metrics (e.g., demographic parity or equalized odds) are evaluated on the same distribution used for pruning. This leaves open whether the extracted subnetwork generalizes to a shifted test distribution whose bias statistics differ from the training set—an essential check for the “bias-agnostic” claim.
minor comments (2)
  1. [§3] Notation for the binary mask m is introduced without a clear statement of its cardinality or how the pruning ratio is chosen; a short paragraph or table entry would clarify reproducibility.
  2. [Figure 2] Figure 2 caption does not indicate whether error bars are standard deviation across seeds or across datasets; this affects interpretation of the performance gap.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and outline revisions to improve clarity and strengthen the experimental validation of the bias-agnostic claims.

read point-by-point responses

  1. Referee: [§3] §3 (BISE algorithm): the mask-selection objective is not shown to be free of an external bias signal. The description states that pruning uses only the original training loss, yet the selection must still quantify “reliance on biased features.” Without an explicit equation or pseudocode step that defines this quantification (e.g., a gradient-based attribution term or a fairness regularizer), it is impossible to verify that the procedure avoids the very information it claims to forgo.

    Authors: We appreciate this observation. The BISE mask-selection procedure optimizes solely the standard cross-entropy loss on the original (biased) training data, without any fairness regularizer, gradient attribution to sensitive attributes, or external bias signal. Subnetworks are identified via iterative magnitude-based pruning of parameters that contribute least to loss minimization, consistent with lottery-ticket hypotheses. To eliminate ambiguity, we will add an explicit mathematical formulation of the mask objective (Equation X) and pseudocode in §3, confirming that no bias-related term enters the selection process. Fairness metrics are computed only after extraction for evaluation purposes. revision: yes

  2. Referee: [§4.2] §4.2 (experimental protocol): the reported fairness metrics (e.g., demographic parity or equalized odds) are evaluated on the same distribution used for pruning. This leaves open whether the extracted subnetwork generalizes to a shifted test distribution whose bias statistics differ from the training set—an essential check for the “bias-agnostic” claim.

    Authors: We agree that demonstrating robustness to distribution shifts in bias statistics is necessary to fully support the bias-agnostic claim. Our current protocol follows standard benchmark splits, but we will augment §4.2 with additional experiments on synthetically shifted test sets (e.g., by varying the strength of spurious correlations between protected attributes and labels while keeping the training distribution fixed). Updated tables and figures will report accuracy and fairness metrics under these conditions. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical pruning strategy with no self-referential derivations

full rationale

The paper introduces BISE as a pruning-based method to isolate existing bias-free subnetworks from vanilla-trained models without retraining or extra unbiased data. No equations, parameter fits presented as predictions, or self-citation chains appear in the abstract or description. The central claim rests on experimental results on benchmarks rather than any derivation that reduces by construction to its inputs. The selection criterion is described at a high level as identifying subnetworks that rely less on biased features, but without shown mathematical reduction or load-bearing self-citation, the approach remains self-contained and non-circular.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities; the core premise that bias-free subnetworks pre-exist in vanilla models is treated as an unstated domain assumption.

pith-pipeline@v0.9.0 · 5503 in / 939 out tokens · 46042 ms · 2026-05-15T16:09:50.791851+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

114 extracted references · 114 canonical work pages · 2 internal anchors

  1. [1]

    The EU artificial intelligence act, 2024

    EU Artificial Intelligence Act. The EU artificial intelligence act, 2024. 2

    work page 2024

  2. [2]

    Does data repair lead to fair models? curating con- textually fair data to reduce model bias

    Sharat Agarwal, Sumanyu Muku, Saket Anand, and Chetan Arora. Does data repair lead to fair models? curating con- textually fair data to reduce model bias. InWACV, 2022. 2

    work page 2022

  3. [3]

    Systematic generalisation with group in- variant predictions

    Faruk Ahmed, Yoshua Bengio, Harm Van Seijen, and Aaron Courville. Systematic generalisation with group in- variant predictions. InICLR, 2021. 7

    work page 2021

  4. [4]

    Mind the gap: Challenges of deep learning approaches to theory of mind.Artificial Intelligence Review, 2023

    Jaan Aru, Aqeel Labash, Oriol Corcoll, and Raul Vicente. Mind the gap: Challenges of deep learning approaches to theory of mind.Artificial Intelligence Review, 2023. 1

    work page 2023

  5. [5]

    Learning de-biased represen- tations with biased representations

    Hyojin Bahng, Sanghyuk Chun, Sangdoo Yun, Jaegul Choo, and Seong Joon Oh. Learning de-biased represen- tations with biased representations. InICML, 2020. 2, 3, 5, 6, 13, 14, 15

    work page 2020

  6. [6]

    Un- biased supervised contrastive learning

    Carlo Alberto Barbano, Benoit Dufumier, Enzo Tartaglione, Marco Grangetto, and Pietro Gori. Un- biased supervised contrastive learning. InICLR, 2023. 2, 16

    work page 2023

  7. [7]

    Unsupervised learning of unbiased visual repre- sentations.IEEE TAI, 2025

    Carlo Alberto Barbano, Enzo Tartaglione, and Marco Grangetto. Unsupervised learning of unbiased visual repre- sentations.IEEE TAI, 2025. 3

    work page 2025

  8. [8]

    Venkatesh Babu

    Abhipsa Basu, Saswat Subhajyoti Mallick, and R. Venkatesh Babu. Mitigating biases in blackbox feature extractors for image classification tasks. In NeurIPS, 2024. 2

    work page 2024

  9. [9]

    Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation

    Yoshua Bengio, Nicholas L ´eonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation.arXiv preprint arXiv:1308.3432, 2013. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  10. [10]

    Simon says: Evaluating and mitigating bias in pruned neural networks with knowledge distillation.arXiv preprint arXiv:2106.07849, 2021

    Cody Blakeney, Nathaniel Huish, Yan Yan, and Ziliang Zong. Simon says: Evaluating and mitigating bias in pruned neural networks with knowledge distillation.arXiv preprint arXiv:2106.07849, 2021. 2

    work page arXiv 2021

  11. [11]

    Nuanced metrics for measur- ing unintended bias with real data for text classification

    Daniel Borkan, Lucas Dixon, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. Nuanced metrics for measur- ing unintended bias with real data for text classification. In Companion proceedings of the 2019 world wide web con- ference, 2019. 5, 15

    work page 2019

  12. [12]

    Simplify: A python library for optimizing pruned neural networks

    Andrea Bragagnolo and Carlo Alberto Barbano. Simplify: A python library for optimizing pruned neural networks. SoftwareX, 2022. 16

    work page 2022

  13. [13]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, Sand- hini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, S...

    work page 2020

  14. [14]

    Rubi: Reducing unimodal biases for visual question answering

    Remi Cadene, Corentin Dancette, Matthieu Cord, and Devi Parikh. Rubi: Reducing unimodal biases for visual question answering. InNeurIPS, 2019. 3, 6

    work page 2019

  15. [15]

    Fairness without demographics through knowledge distillation

    Junyi Chai, Taeuk Jang, and Xiaoqian Wang. Fairness without demographics through knowledge distillation. In NeurIPS, 2022. 2

    work page 2022

  16. [16]

    Knowledge distillation with the reused teacher classifier

    Defang Chen, Jian-Ping Mei, Hailin Zhang, Can Wang, Yan Feng, and Chun Chen. Knowledge distillation with the reused teacher classifier. InCVPR, 2022. 7

    work page 2022

  17. [17]

    A survey on deep neural network pruning: Taxonomy, com- parison, analysis, and recommendations.IEEE TPAMI,

    Hongrong Cheng, Miao Zhang, and Javen Qinfeng Shi. A survey on deep neural network pruning: Taxonomy, com- parison, analysis, and recommendations.IEEE TPAMI,

    work page

  18. [18]

    Don’t take the easy way out: Ensemble based methods for avoiding known dataset biases.EMNLP-IJCNLP, 2019

    Christopher Clark, Mark Yatskar, and Luke Zettlemoyer. Don’t take the easy way out: Ensemble based methods for avoiding known dataset biases.EMNLP-IJCNLP, 2019. 6

    work page 2019

  19. [19]

    Environment inference for invariant learning

    Elliot Creager, J ¨orn-Henrik Jacobsen, and Richard Zemel. Environment inference for invariant learning. InICML,

    work page

  20. [20]

    Conscientious classification: A data scientist’s guide to discrimination-aware classification.Big data, 2017

    Brian d’Alessandro, Cathy O’Neil, and Tom LaGatta. Conscientious classification: A data scientist’s guide to discrimination-aware classification.Big data, 2017. 2

    work page 2017

  21. [21]

    Fairdistillation: Mit- igating stereotyping in language models.ECML-PKDD,

    Pieter Delobelle and Bettina Berendt. Fairdistillation: Mit- igating stereotyping in language models.ECML-PKDD,

    work page

  22. [22]

    Bert: Pre-training of deep bidirectional trans- formers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional trans- formers for language understanding. InProceedings of the 2019 conference of the North American chapter of the as- sociation for computational linguistics: human language technologies, volume 1 (long and short papers), 2019. 1, 5, 16

    work page 2019

  23. [23]

    A winning hand: Com- pressing deep networks can improve out-of-distribution ro- bustness

    James Diffenderfer, Brian Bartoldson, Shreya Chaganti, Jize Zhang, and Bhavya Kailkhura. A winning hand: Com- pressing deep networks can improve out-of-distribution ro- bustness. InNeurIPS, 2021. 3

    work page 2021

  24. [24]

    The lottery ticket hypothesis: Finding sparse, trainable neural networks

    Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In ICLR, 2019. 3

    work page 2019

  25. [25]

    Shortcut learning in deep neural net- works.Nature Machine Intelligence, 2020

    Robert Geirhos, J ¨orn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Fe- lix A Wichmann. Shortcut learning in deep neural net- works.Nature Machine Intelligence, 2020. 1, 2, 4

    work page 2020

  26. [26]

    Debias- ing pre-trained language models via efficient fine-tuning

    Michael Gira, Ruisu Zhang, and Kangwook Lee. Debias- ing pre-trained language models via efficient fine-tuning. In Proceedings of the second workshop on language technol- ogy for equality, diversity and inclusion, 2022. 2

    work page 2022

  27. [27]

    Morphnet: Fast & simple resource-constrained structure learning of deep networks

    Ariel Gordon, Elad Eban, Ofir Nachum, Bo Chen, Hao Wu, Tien-Ju Yang, and Edward Choi. Morphnet: Fast & simple resource-constrained structure learning of deep networks. InCVPR, 2018. 3

    work page 2018

  28. [28]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InCVPR,

    work page

  29. [29]

    Structured pruning for deep convolutional neural networks: A survey.IEEE TPAMI,

    Yang He and Lingao Xiao. Structured pruning for deep convolutional neural networks: A survey.IEEE TPAMI,

    work page

  30. [30]

    Benchmarking neural network robustness to common corruptions and per- turbations

    Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and per- turbations. InICLR, 2019. 5, 13, 15

    work page 2019

  31. [31]

    Unbiased classifica- tion through bias-contrastive and bias-balanced learning

    Youngkyu Hong and Eunho Yang. Unbiased classifica- tion through bias-contrastive and bias-balanced learning. In NeurIPS, 2021. 2, 6, 7, 13, 18, 20

    work page 2021

  32. [32]

    What do compressed deep neural networks forget?arXiv preprint arXiv:1911.05248,

    Sara Hooker, Aaron Courville, Gregory Clark, Yann Dauphin, and Andrea Frome. What do compressed deep neural networks forget?arXiv preprint arXiv:1911.05248,

    work page arXiv 1911

  33. [33]

    Charac- terising bias in compressed models.arXiv preprint arXiv:2010.03058,

    Sara Hooker, Nyalleng Moorosi, Gregory Clark, Samy Bengio, and Emily Denton. Characterising bias in com- pressed models.arXiv preprint arXiv:2010.03058, 2020. 3

    work page arXiv 2010

  34. [34]

    Selecmix: Debiased learning by contradicting-pair sam- pling

    Inwoo Hwang, Sangjun Lee, Yunhyeok Kwak, Seong Joon Oh, Damien Teney, Jin-Hwa Kim, and Byoung-Tak Zhang. Selecmix: Debiased learning by contradicting-pair sam- pling. InNeurIPS, 2022. 2

    work page 2022

  35. [35]

    Simple data balancing achieves com- petitive worst-group-accuracy

    Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves com- petitive worst-group-accuracy. InProceedings of the First Conference on Causal Learning and Reasoning. PMLR,

    work page

  36. [36]

    Bias in pruned vision models: In-depth analysis and counter- measures

    Eugenia Iofinova, Alexandra Peste, and Dan Alistarh. Bias in pruned vision models: In-depth analysis and counter- measures. InCVPR, 2023. 3

    work page 2023

  37. [37]

    On feature learning in the presence of spu- rious correlations

    Pavel Izmailov, Polina Kirichenko, Nate Gruver, and An- drew G Wilson. On feature learning in the presence of spu- rious correlations. InNeurIPS, 2022. 2, 5, 7, 15, 16

    work page 2022

  38. [38]

    On the effect of pruning on adversarial robustness

    Artur Jord ˜ao and H´elio Pedrini. On the effect of pruning on adversarial robustness. InICCV Workshops, 2021. 3

    work page 2021

  39. [39]

    Going be- yond classification accuracy metrics in model compression

    Vinu Joseph, Shoaib Ahmed Siddiqui, Aditya Bhaskara, Ganesh Gopalakrishnan, Saurav Muralidharan, Michael Garland, Sheraz Ahmed, and Andreas Dengel. Going be- yond classification accuracy metrics in model compression. arXiv preprint arXiv:2012.01604, 2020. 3

    work page arXiv 2012

  40. [40]

    Learning not to learn: Training deep neu- ral networks with biased data

    Byungju Kim, Hyunwoo Kim, Kyungsu Kim, Sungjin Kim, and Junmo Kim. Learning not to learn: Training deep neu- ral networks with biased data. InCVPR, 2019. 2, 3, 7

    work page 2019

  41. [41]

    Learning debiased classifier with biased committee

    Nayeong Kim, Sehyun Hwang, Sungsoo Ahn, Jaesik Park, and Suha Kwak. Learning debiased classifier with biased committee. InNeurIPS, 2022. 2, 3, 7, 8

    work page 2022

  42. [42]

    Improving robustness to multiple spuri- ous correlations by multi-objective optimization

    Nayeong Kim, Juwon Kang, Sungsoo Ahn, Jungseul Ok, and Suha Kwak. Improving robustness to multiple spuri- ous correlations by multi-objective optimization. InICML,

    work page

  43. [43]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InICLR, 2015. 16

    work page 2015

  44. [44]

    Wilds: A benchmark of in-the-wild distribution shifts

    Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, Tony Lee, Etienne David, Ian Stavness, Wei Guo, Berton Earnshaw, Imran Haque, Sara M Beery, Jure Leskovec, Anshul Kundaje, Emma Pierson, Sergey Levine, Chelsea Finn, and Percy Liang. Wilds: A ...

    work page 2021

  45. [45]

    Achieving fairness through channel pruning for dermatological disease diagnosis

    Qingpeng Kong, Ching-Hao Chiu, Dewen Zeng, Yu-Jen Chen, Tsung-Yi Ho, Jingtong Hu, and Yiyu Shi. Achieving fairness through channel pruning for dermatological disease diagnosis. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, 2024. 3

    work page 2024

  46. [46]

    Algorithmic bias: review, synthesis, and future research directions.Eu- ropean Journal of Information Systems, 2022

    Nima Kordzadeh and Maryam Ghasemaghaei. Algorithmic bias: review, synthesis, and future research directions.Eu- ropean Journal of Information Systems, 2022. 1

    work page 2022

  47. [47]

    A contrastive learning approach to mitigate bias in speech models.arXiv preprint arXiv:2406.14686,

    Alkis Koudounas, Flavio Giobergia, Eliana Pastor, and Elena Baralis. A contrastive learning approach to mitigate bias in speech models.arXiv preprint arXiv:2406.14686,

    work page arXiv

  48. [48]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. Master’s thesis, Department of Computer Sci- ence, University of Toronto, 2009. 5, 15

    work page 2009

  49. [49]

    Gradient-based learning applied to document recognition.Proceedings of the IEEE, 1998

    Yann LeCun, L ´eon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 1998. 5, 14

    work page 1998

  50. [50]

    Debiased distillation by trans- planting the last layer.arXiv preprint arXiv:2302.11187,

    Jiwoon Lee and Jaeho Lee. Debiased distillation by trans- planting the last layer.arXiv preprint arXiv:2302.11187,

    work page arXiv

  51. [51]

    Learning debiased representation via dis- entangled feature augmentation

    Jungsoo Lee, Eungyeup Kim, Juyoung Lee, Jihyeon Lee, and Jaegul Choo. Learning debiased representation via dis- entangled feature augmentation. InNeurIPS, 2021. 2, 6, 20

    work page 2021

  52. [52]

    Repair: Removing represen- tation bias by dataset resampling

    Yi Li and Nuno Vasconcelos. Repair: Removing represen- tation bias by dataset resampling. InCVPR, 2019. 2

    work page 2019

  53. [53]

    Discover and mitigate unknown biases with debiasing alternate net- works

    Zhiheng Li, Anthony Hoogs, and Chenliang Xu. Discover and mitigate unknown biases with debiasing alternate net- works. InECCV, 2022. 2, 5, 6, 7, 8, 14, 15, 16

    work page 2022

  54. [54]

    Achieving adversarial ro- bustness via sparsity.Machine Learning, 2022

    Ningyi Liao, Shufan Wang, Liyao Xiang, Nanyang Ye, Shuo Shao, and Pengzhi Chu. Achieving adversarial ro- bustness via sparsity.Machine Learning, 2022. 3

    work page 2022

  55. [55]

    Lost in pruning: The effects of pruning neural networks beyond test accuracy.Proceedings of Machine Learning and Systems, 2021

    Lucas Liebenwein, Cenk Baykal, Brandon Carter, David Gifford, and Daniela Rus. Lost in pruning: The effects of pruning neural networks beyond test accuracy.Proceedings of Machine Learning and Systems, 2021. 3

    work page 2021

  56. [56]

    Biasadv: Bias-adversarial augmentation for model debiasing

    Jongin Lim, Youngdong Kim, Byungjai Kim, Chanho Ahn, Jinwoo Shin, Eunho Yang, and Seungju Han. Biasadv: Bias-adversarial augmentation for model debiasing. In CVPR, 2023. 2

    work page 2023

  57. [57]

    Liu, Behzad Haghgoo, Annie S

    Evan Z. Liu, Behzad Haghgoo, Annie S. Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robust- ness without training group information. InICML, 2021. 2, 7, 8

    work page 2021

  58. [58]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. InICCV, 2015. 5, 13, 14, 15

    work page 2015

  59. [59]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InICLR, 2019. 16

    work page 2019

  60. [60]

    Conditional contrastive learning for im- proving fairness in self-supervised learning.arXiv preprint arXiv:2106.02866, 2021

    Martin Q Ma, Yao-Hung Hubert Tsai, Paul Pu Liang, Han Zhao, Kun Zhang, Ruslan Salakhutdinov, and Louis- Philippe Morency. Conditional contrastive learning for im- proving fairness in self-supervised learning.arXiv preprint arXiv:2106.02866, 2021. 2 10

    work page arXiv 2021

  61. [61]

    An image enhancing pattern- based sparsity for real-time inference on mobile devices

    Xiaolong Ma, Wei Niu, Tianyun Zhang, Sijia Liu, Sheng Lin, Hongjia Li, Xiang Chen, Jian Tang, Kaisheng Ma, Bin Ren, and Yanzhi Wang. An image enhancing pattern- based sparsity for real-time inference on mobile devices. In ECCV, 2020. 3

    work page 2020

  62. [62]

    De- biasing deep chest x-ray classifiers using intra-and post- processing methods

    Ricards Marcinkevics, Ece Ozkan, and Julia E V ogt. De- biasing deep chest x-ray classifiers using intra-and post- processing methods. InMachine Learning for Healthcare Conference. PMLR, 2022. 2

    work page 2022

  63. [63]

    A survey on bias and fairness in machine learning.ACM computing surveys (CSUR), 2021

    Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning.ACM computing surveys (CSUR), 2021. 2

    work page 2021

  64. [64]

    Debiasing masks: A new framework for shortcut mitigation in NLU.EMNLP, 2022

    Johannes Mario Meissner, Saku Sugawara, and Akiko Aizawa. Debiasing masks: A new framework for shortcut mitigation in NLU.EMNLP, 2022. 2, 3

    work page 2022

  65. [65]

    Gender arti- facts in visual datasets

    Nicole Meister, Dora Zhao, Angelina Wang, Vikram V Ra- maswamy, Ruth Fong, and Olga Russakovsky. Gender arti- facts in visual datasets. InICCV, 2023. 1

    work page 2023

  66. [66]

    Pruning filter in filter

    Fanxu Meng, Hao Cheng, Ke Li, Huixiang Luo, Xiaowei Guo, Guangming Lu, and Xing Sun. Pruning filter in filter. InNeurIPS, 2020. 3

    work page 2020

  67. [67]

    A fair loss function for network pruning.NeurIPS TSRML Workshop, 2022

    Robbie Meyer and Alexander Wong. A fair loss function for network pruning.NeurIPS TSRML Workshop, 2022. 3

    work page 2022

  68. [68]

    Mining bias-target alignment from voronoi cells

    R ´emi Nahon, Van-Tam Nguyen, and Enzo Tartaglione. Mining bias-target alignment from voronoi cells. InICCV,

    work page

  69. [69]

    Debiasing surgeon: fan- tastic weights and how to find them

    R ´emi Nahon, Ivan Luiz De Moura Matos, Van-Tam Nguyen, and Enzo Tartaglione. Debiasing surgeon: fan- tastic weights and how to find them. InECCV, 2024. 2, 3, 7, 8, 14, 16, 17

    work page 2024

  70. [70]

    Learning from failure: De-biasing classifier from biased classifier

    Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: De-biasing classifier from biased classifier. InNeurIPS, 2020. 2, 3, 4, 5, 6, 7, 8, 15, 16, 18

    work page 2020

  71. [71]

    Efficient adaptation of deep neural networks for semantic segmentation in space applications.Scientific Reports, 2025

    Leonardo Olivi, Edoardo Santero Mormile, and Enzo Tartaglione. Efficient adaptation of deep neural networks for semantic segmentation in space applications.Scientific Reports, 2025. 17

    work page 2025

  72. [72]

    Prune responsibly.arXiv preprint arXiv:2009.09936, 2020

    Michela Paganini. Prune responsibly.arXiv preprint arXiv:2009.09936, 2020. 3

    work page arXiv 2009

  73. [73]

    Training debiased subnetworks with con- trastive weight pruning

    Geon Yeong Park, Sangmin Lee, Sang Wan Lee, and Jong Chul Ye. Training debiased subnetworks with con- trastive weight pruning. InCVPR, 2023. 2

    work page 2023

  74. [74]

    Self-supervised debi- asing using low rank regularization

    Geon Yeong Park, Chanyong Jung, Sangmin Lee, Jong Chul Ye, and Sang Wan Lee. Self-supervised debi- asing using low rank regularization. InCVPR, 2024. 6, 13

    work page 2024

  75. [75]

    More, Christian M

    Ot ´avio Parraga, Martin D. More, Christian M. Oliveira, Nathan S. Gavenski, Lucas S. Kupssinsk ¨u, Adilson Medronha, Luis V . Moura, Gabriel S. Sim ˜oes, and Ro- drigo C. Barros. Debiasing methods for fairer neural mod- els in vision and language research: A survey.ACM com- puting surveys (CSUR), 2022. 2

    work page 2022

  76. [76]

    Learning deep representations with probabilistic knowledge transfer

    Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 2

    work page 2018

  77. [77]

    Looking at model debiasing through the lens of anomaly detection

    Vito Paolo Pastore, Massimiliano Ciranni, Davide Marinelli, Francesca Odone, and Vittorio Murino. Looking at model debiasing through the lens of anomaly detection. InWACV, 2025. 2

    work page 2025

  78. [78]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Rai- son, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-per...

    work page 2019

  79. [79]

    Learning unbiased representations via mu- tual information backpropagation

    Ruggero Ragonesi, Riccardo V olpi, Jacopo Cavazza, and Vittorio Murino. Learning unbiased representations via mu- tual information backpropagation. InCVPR, 2021. 2

    work page 2021

  80. [80]

    A comparative study on the impact of model compression techniques on fairness in language models

    Krithika Ramesh, Arnav Chavan, Shrey Pandit, and Sunayana Sitaram. A comparative study on the impact of model compression techniques on fairness in language models. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2023. 3

    work page 2023

Showing first 80 references.