[Paper Review] Towards Fairness in Visual Recognition: Effective Strategies for Bias Mitigation

Zeyu Wang, Klint Qinami+5 more|arXiv (Cornell University)|Nov 26, 2019

Domain Adaptation and Few-Shot Learning55 references21 citations

TL;DR

This paper introduces a controlled benchmark to evaluate bias mitigation in visual recognition, demonstrating that a simple domain-independent training strategy outperforms complex adversarial methods. It shows that the proposed approach reduces gender bias in attribute classification on CelebA while improving mAP from 74.7% to 76.3%.

ABSTRACT

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.

Motivation & Objective

• To address the lack of systematic comparison among bias mitigation techniques in deep visual recognition models.
• To design a controlled benchmark that isolates and quantifies spurious correlations (e.g., gender-attribute skew) in training data.
• To evaluate and compare existing mitigation strategies—adversarial training, bias amplification reduction, domain-specific learning—under identical conditions.
• To validate findings on real-world data, specifically gender bias in the CelebA attribute classification task.
• To propose and demonstrate a domain-independent training method that effectively mitigates bias without requiring domain-specific adaptation.

Proposed method

• Construct a synthetic visual recognition benchmark using dataset augmentation to introduce controlled, spurious correlations between input attributes (e.g., gender) and labels (e.g., activity).
• Train a two-branch CNN with shared features and separate prediction heads for each domain (e.g., male/female), enabling domain-specific and domain-independent inference.
• Implement a domain-independent inference strategy that computes the prediction as the weighted average of domain-specific outputs, using domain-conditional probabilities.
• Use a multi-task loss combining classification loss and a domain classifier loss to encourage disentanglement of sensitive attributes from task-relevant features.
• Apply inference-time correction via the RBA (Reducing Bias Amplification) method of Zhao et al. for comparison.
• Evaluate models using mAP (mean average precision) and bias amplification (BA) metrics, with BA measuring the correlation between gender and attribute predictions.

Experimental results

Research questions

• RQ1 How do different bias mitigation techniques perform under controlled, isolated spurious correlations in visual recognition?
• RQ2 Can adversarial training effectively reduce bias without sacrificing accuracy in visual recognition tasks?
• RQ3 Does a domain-independent training strategy outperform domain-specific or adversarial approaches in mitigating bias?
• RQ4 How does the level of gender imbalance in attributes affect the performance gain of the proposed method on real-world data?
• RQ5 Can the proposed method generalize to multi-label, real-world benchmarks like CelebA while maintaining high accuracy and fairness?

Key findings

• Adversarial training reduces bias amplification (BA from 0.010 to 0.019) but harms mAP (71.9% vs. 74.7% for baseline), indicating a trade-off between fairness and accuracy.
• The domain-independent training method achieves the highest mAP (76.3%) and the lowest bias amplification (-0.035), outperforming both baseline and adversarial baselines.
• The domain-independent model improves mAP by 1.6 percentage points over the baseline on the CelebA test set, with the largest gains on attributes with high gender imbalance (e.g., >80% skew).
• The correlation between attribute skew and performance gain is strong (ρ = 0.709), confirming that the method is most effective when bias is pronounced.
• Oversampling achieves the highest mAP (77.6%) but is less effective in high-skew, low-similarity settings, highlighting trade-offs between data augmentation and architectural design.
• The proposed domain-independent method is simpler and more effective than complex adversarial or inference-time correction methods, especially in real-world, multi-label settings.