Learning Adversarially Fair and Transferable Representations
By David Madras, Elliot Creager, Toniann Pitassi, Richard S. Zemel. 2018
796 citations recorded by the source corpus.
In this paper, we advocate for representation learning as the key to mitigating unfair prediction outcomes downstream. Motivated by a scenario where learned representations are used by third parties with unknown objectives, we propose and explore adversarial representation learning as a natural method of ensuring those parties act fairly. We connect group fairness (demographic parity, equalized odds, and equal opportunity) to different adversarial objectives. Through worst-case theoretical guarantees and experimental validation, we show that the choice of this objective is crucial to fair prediction. Furthermore, we present the first in-depth experimental demonstration of fair transfer learning and demonstrate empirically that our learned representations admit fair predictions on new tasks while maintaining utility, an essential goal of fair representation learning.
Subfields: Artificial Intelligence, Safety Research, Representation & Self-Supervised Learning, AI Ethics, Fairness & Society
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Yoshua Bengio
- Percy Liang
- Dacheng Tao
- Bernhard Schölkopf
- Mohit Bansal
- Ruslan Salakhutdinov
- Sergey Levine
- Quoc V. Le
In Safety Research
- Silvia Chiappa
- Dana Pessach
- Erez Shmueli
- Luca Oneto
- Krishna P. Gummadi
- Isabel Valera
- Lu Zhang
- Muhammad Bilal Zafar
In Representation & Self-Supervised Learning
- Bernhard Schölkopf
- Dacheng Tao
- William L. Hamilton
- Francesco Locatello
- Stefano Ermon
- Stefan Bauer
- Olivier Bachem
- Yoshua Bengio
In AI Ethics, Fairness & Society
- Silvia Chiappa
- Luca Oneto
- Dana Pessach
- Erez Shmueli
- Christian Haas
- Simon Caton
- Krishna P. Gummadi
- Fred Morstatter
Citations in the corpus
- A Step Toward More Inclusive People Annotations for Fairness (2021, background): “Finally, many bias mitigation strategies require attribute labels on at least a portion of the model’s training set [3, 19].”
- Towards fairer datasets: filtering and balancing the distribution of the people subtree in the ImageNet hierarchy (2019, background): “Algorithmic solutions have ranged from removing undesired bias by preprocessing the data [39, 59], striking a tradeoff between performance and fairness by posing additional regularization during training or inference [34, 40, 52, 80, 81, 83], or designing application-specific interventions (such as ”
- Attributing Fair Decisions with Attention Interventions (2021, methodology): “In addition to informationtheoretic approaches, we also considered baselines that use adversarial learning such as MaxEnt-ARL (Roy and Boddeti 2019), LAFTR (Madras et al. 2018), and Adversarial Forgetting (Jaiswal et al. 2020).”
- Compositional Fairness Constraints for Graph Embeddings (2019, methodology): “Our experimental setup closely mirrors that of (Madras et al., 2018) where we jointly train the main model with adversaries, but when testing invariance, we train a new classifier (with the same capacity as the discriminator) to predict the senstive attributes from the learned embeddings.”
- The Price of Fair PCA: One Extra Dimension (2018, background): “The first piece is a broad line of work [Zemel et al., 2013; Beutel et al., 2017; Calmon et al., 2017; Madras et al., 2018; Zhang et al., 2018] that aims to design representations which will be conditionally independent of the protected attribute, while retaining as much information as possible (and…”
- FairGAN: Fairness-aware Generative Adversarial Networks (2018, background): “Studies in [20], [21] further proposed various adversarial objectives to achieve different levels of group fairness including demographic parity, equalized odds and equal opportunity.”
- On the Fairness of Disentangled Representations (2019, background): “Furthermore, there are several approaches that employ adversarial [30] training to avoid information leakage between the sensitive attributes and the representation [23, 64, 96].”
- Learning Fair Rule Lists (2019, methodology): “For LAFTR, we follow a similar validation procedure as the authors (Madras et al., 2018).”
- Measuring and Reducing Gendered Correlations in Pre-trained Models (2020, neutral): “Another popular technique uses adversarial losses in order to remove demographic information from learned representations (Louizos et al., 2015; Edwards and Storkey, 2015; Beutel et al., 2017; Zhang et al., 2018; Elazar and Goldberg, 2018; Madras et al., 2018).”
- Learning Controllable Fair Representations (2018, methodology): “Our work have several notable differences from prior methods (such as LAFTR (Madras et al., 2018)) that make it hard to compare them directly.”
- Algorithmic Fairness (2020, background): “…also note that learning fairly to predict the outcome of an unprivileged group can be thought of as learning to predict in a different domain [48, 96], and therefore, notions from the field of domain adaptation can be adopted to enhance the study of the field of algorithmic fairness , and vice…”
- On Transferability of Bias Mitigation Effects in Language Model Fine-Tuning (2021, methodology): “We implement UBM with two different bias mitigation algorithms in the upstream bias mitigation phase: explanation regularization (Kennedy et al., 2020), and adversarial de-biasing (Zhang et al., 2018; Madras et al., 2018; Xia et al., 2020 where α is a trade-off hyperparameter.”
- Censored and Fair Universal Representations using Generative Adversarial Models (2019, methodology): “Note that ∆EO is the measure of equalized odds defined in [Madras et al., 2018] and ∆EO(> 50K) + ∆EO(≤ 50K) = ∆EO”
- A Review on Fairness in Machine Learning (2022, background): “distinguish which samples belong to a privileged group and which belong to an unprivileged group [23, 66, 141].”