Learning Unbiased Representations via Mutual Information Backpropagation

Ragonesi, Ruggero; Volpi, R.; Cavazza, Jacopo; Murino, Vittorio

doi:10.48550/arxiv.2003.06430

Cited by 1 publication

(4 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Training σ = 0.020 σ= 0.025 σ = 0.030 σ = 0.035 σ = 0.040 σ = 0.045 σ = 0.050 ERM (λ = 0.0) 0.476 ± 0.005 0.542 ± 0.004 0.664 ± 0.001 0.720 ± 0.010 0.785 ± 0.003 0.838 ± 0.002 0.870 ± 0.001 Ragonesi [38] 0.592 ± 0.018 0.678 ± 0.015 0.737 ± 0.028 0.795 ± 0.012 0.814 ± 0.019 0.837 ± 0.004 0.877 ± 0.010 Zhang et al [50] 0.584 ± 0.034 0.625 ± 0.033 0.709 ± 0.027 0.733 ± 0.020 0.807 ± 0.013 0.803 ± 0.027 0.831 ± 0.027 Kim et Our experiments on real-world data are performed on five data sets. In three data sets, the sensitive and the outcome true value are both continuous: the US Census data set [44], the Motor data set [42] and the Crime data set [15].…”

Section: Color Variancementioning

confidence: 99%

“…Adel et al [1] learn a fair representation by inputting it to an adversary network, which is prevented from predicting the sensitive attribute (upper left in Figure 1 in appendix). Other papers minimize the mutual information between the representation and the sensitive attribute: Kim et al [26] rely on adversarial training with a discriminator detecting the bias, while Ragonesi et al [38] rely on an estimation by neural network of mutual information [6] (lower left in Figure 1 in appendix).…”

Section: Related Workmentioning

confidence: 99%

“…However, this attenuation is performed on the predictor output only. Several recent papers [1,38] have shown that performing the attenuation on a representation tends to give better results in terms of prediction accuracy while remaining fair in complex real-world scenarios. In this work, we are interested in learning fair representations via this Renyi estimator.…”

Section: Metrics For Continuous Statistical Dependencementioning

confidence: 99%

“…Grari et al [19] minimize the HGR correlation between the prediction output and the sensitive attribute (middle right). Ragonesi et al [38] rely on MINE to minimize the mutual information between a representation and the sensitive attribute (lower left). Another approach, referred to as Fair MINE NN, minimizes the mutual information between the prediction output and the sensitive attribute (lower right).…”

Section: Fair Adversarial Architecturesmentioning

confidence: 99%

See 3 more Smart Citations

Learning Unbiased Representations via Rényi Minimization

Grari¹,

Hajouji²,

Lamprier³

et al. 2020

Preprint

View full text Add to dashboard Cite

In recent years, significant work has been done to include fairness constraints in the training objective of machine learning algorithms. Many state-of the-art algorithms tackle this challenge by learning a fair representation which captures all the relevant information to predict the output Y while not containing any information about a sensitive attribute S. In this paper, we propose an adversarial algorithm to learn unbiased representations via the Hirschfeld-Gebelein-Renyi (HGR) maximal correlation coefficient. We leverage recent work which has been done to estimate this coefficient by learning deep neural network transformations and use it as a minmax game to penalize the intrinsic bias in a multi dimensional latent representation. Compared to other dependence measures, the HGR coefficient captures more information about the non-linear dependencies with the sensitive variable, making the algorithm more efficient in mitigating bias in the representation. We empirically evaluate and compare our approach and demonstrate significant improvements over existing works in the field.Preprint. Under review.

show abstract

Section: Color Variancementioning

confidence: 99%