Towards Fairness in Visual Recognition: Effective Strategies for Bias Mitigation

By Zeyu Wang, Klint Qinami, Yannis Karakozis, Kyle Genova, Prem Qu Nair, Kenji Hata, Olga Russakovsky. 2019

430 citations recorded by the source corpus.

Computer vision models learn to perform a task by capturing relevant statistics from training data. It has been shown that models learn spurious age, gender, and race correlations when trained for seemingly unrelated tasks like activity recognition or image captioning. Various mitigation techniques have been presented to prevent models from utilizing or learning such biases. However, there has been little systematic comparison between these techniques. We design a simple but surprisingly effective visual recognition benchmark for studying bias mitigation. Using this benchmark, we provide a thorough analysis of a wide range of techniques. We highlight the shortcomings of popular adversarial training approaches for bias mitigation, propose a simple but similarly effective alternative to the inference-time Reducing Bias Amplification method of Zhao et al., and design a domain-independent training technique that outperforms all other methods. Finally, we validate our findings on the attribute classification task in the CelebA dataset, where attribute presence is known to be correlated with the gender of people in the image, and demonstrate that the proposed technique is effective at mitigating real-world gender bias.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

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

In Computer Vision and Pattern Recognition

Citations in the corpus

DISCO: Dynamic and Invariant Sensitive Channel Obfuscation for deep neural networks (2020, background): “A popular category of techniques involve adversarial representation learning [34, 29, 35] to mitigate the impact of the bias attribute on the task attribute.”
Quantifying Societal Bias Amplification in Image Captioning (2022, methodology): “In addition, REVISE [37] and CIFAR-10S [41] ease the task of identifying bias on datasets and models, respectively.”
Handling Bias in Toxic Speech Detection: A Survey (2022, background): “For example a benchmark [107] was recently developed to systematically compare gender-bias mitigation techniques in visual recognition models.”
Image Representations Learned With Unsupervised Pre-Training Contain Human-like Biases (2020, background): “Further research is needed to determine how architecture choices affect embedded biases and how dataset filtering and balancing techniques might help [75, 74].”
Are My Deep Learning Systems Fair? An Empirical Study of Fixed-Seed Training (2021, methodology): “Seven bias metrics include demographic parity [15] (DP), normalized disparate impact [15] (DI), statistical parity subgroup fairness [46] (SPSF), false positive subgroup fairness [46] (FPSF), equalized odds false positive [24] (EOFP), equalized odds true positive [35] (EOTP), and bias amplification ”
REVISE: A Tool for Measuring and Mitigating Bias in Visual Datasets (2020, background): “There has been signiﬁcant work looking at how to deal with this from the algorithm side (Dwork et al., 2012; Dwork et al., 2017; Khosla et al., 2012; Wang et al., 2020) with varying deﬁnitions of fairness (Gajane and Pech-enizkiy, 2017; Hardt et al., 2016; Kilbertus et al., 2017; Pleiss et al.,…”
Scaling Vision Transformers to 22 Billion Parameters (2023, background): “For example, they can amplify spurious correlations in the training data (Hendricks et al., 2018; Caliskan et al., 2017; Zhao et al., 2017; Wang et al., 2020) and result in error disparities (Zhao et al., 2017; Buolamwini & Gebru, 2018; Deuschel et al., 2020).”
Counterfactual Generative Networks (2021, background): “This assumption is in line with latest related work on visual bias (Goyal et al., 2019; Wang et al., 2020), which considers a strong correlation to be above 95 %.”
PaLI-X: On Scaling up a Multilingual Vision and Language Model (2023, background): “Large models, if left unchecked, have the potential to inflict harm on society – such as amplifying biases [76, 77, 78, 79], causing disparities [78, 80, 81], or encoding narrow cultural perspectives [82, 83].”
Taxonomizing and Measuring Representational Harms: A Look at Image Tagging (2023, background): “…that depict some social groups than they do for images that depict others, ﬁnding performance disparities between genders (Bhargava and Forsyth 2019; Wang et al. 2020), ages (Wang et al. 2020), races (Buolamwini and Gebru 2018; Wilson, Hoffman, and Morgenstern 2019), socioeconomic statuses…”
Fair Attribute Classification through Latent Space De-biasing (2020, methodology): “Table 5: Comparison of our method with weighted and adversarial training from [41].”
A Survey of Deep Learning for Scientific Discovery (2020, background): “A line of active research studies how to debias these representations [4, 241].”