Leveling Down in Computer Vision: Pareto Inefficiencies in Fair Deep Classifiers

Dominik Zietlow, Michael Lohaus, Guha Balakrishnan, Matthäus Kleindessner, Francesco Locatello, Bernhard Schölkopf, Chris Russell

Introduction

In the pursuit of algorithmic fairness, the machine learning community has traditionally viewed the problem as a trade-off: to improve the performance of a disadvantaged group, one must sacrifice a degree of overall accuracy. This paradigm, while often applicable to low-capacity models, fails to account for the unique dynamics of modern deep learning.

In high-capacity computer vision models, standard fairness interventions often result in a phenomenon the authors call "leveling down." Instead of redistributing performance from a majority group to a minority group, these techniques frequently degrade the accuracy of every group involved. This leads to Pareto Inefficiency, where the model becomes "fairer" only because the performance gap shrinks as everyone’s performance declines. This paper argues that fairness in deep learning is fundamentally a generalization problem and proposes a new framework to achieve Pareto improvements.

Fairness as a Generalization Problem

Traditional fairness methods often focus on modifying the loss function or reweighting samples during training. However, deep neural networks possess the capacity to achieve near-zero training error, effectively memorizing the training set. When a model fits the training data perfectly, fairness constraints are "trivially satisfied" during the training phase.

Consequently, fairness violations observed on unseen test data are not a result of training bias, but rather a failure of generalization. By extending the classic bias-variance decomposition to fairness, the authors demonstrate that for high-capacity models, the disparity in performance between groups is dominated by the variance (generalization error) specific to each sub-distribution. To achieve true fairness without leveling down, models must be encouraged to generalize better to underrepresented or difficult-to-learn groups.

The g-SMOTE Framework

To address the generalization gap, the authors move away from static penalties and toward adaptive, data-centric interventions. They propose a two-pronged strategy:

1. Adaptive Sampling

Rather than using a fixed reweighting scheme, the algorithm iteratively monitors the performance of different groups on a held-out validation set. It identifies which group is currently suffering from the highest generalization error and dynamically prioritizes that group for data augmentation in the next training cycle.

2. Generalized SMOTE (g-SMOTE)

To provide the model with the semantic diversity needed to learn disadvantaged groups, the authors introduce g-SMOTE, a GAN-based evolution of the Synthetic Minority Over-sampling Technique. The process involves:

  • Latent Mapping: Using an invertible GAN (such as InvGAN) to map real images from the training set into a structured latent space.
  • Interpolation: Performing linear interpolation between a specific datapoint and its $k$-nearest neighbors within that latent space.
  • Synthesis: Rendering these interpolated points back into the image domain via the GAN generator.

This approach creates new, label-consistent training samples that offer more semantic variety than standard augmentations like flipping or cropping, helping the model learn the underlying features of the minority distribution.

Experimental Evaluation

The authors evaluated their approach using the CelebA dataset, focusing on 13 gender-independent facial attributes (e.g., "Smiling," "Eyeglasses") to ensure the analysis was not confounded by biological correlations.

Key Findings

  • The Failure of Standard Baselines: Most existing fairness methods—including adversarial debiasing and simple reweighting—were found to be Pareto inefficient. In many cases, these methods decreased the accuracy of the worst-performing group compared to a standard "unfair" baseline, effectively acting as a form of noise that hindered the model's overall learning.
  • Pareto Improvements with g-SMOTE: g-SMOTE + Adaptive Sampling was the only method that consistently achieved Pareto improvements. It boosted the accuracy of the most disadvantaged groups without sacrificing the performance of the majority.
  • Quantitative Success: In a 10k image training scenario on CelebA, while other methods saw the "Minimum Group Accuracy" drop by 1-5%, g-SMOTE improved it from 85.68% to 85.93%.
  • Complementary Effects: The gains from g-SMOTE were found to be orthogonal to standard data augmentation techniques. Combining g-SMOTE with RandAugment yielded the highest overall performance across all groups.

Conclusion and Future Work

The research shifts the fairness conversation in computer vision from "how do we penalize the majority?" to "how do we help the minority generalize?" By treating fairness as a generalization challenge, the authors demonstrate that it is possible to build models that are both fairer and more accurate.

However, the approach is not without limitations. The effectiveness of g-SMOTE is intrinsically tied to the quality of the underlying generative model; if the GAN cannot accurately represent the minority group, the synthetic data will be of little use. Additionally, the computational cost of training invertible GANs is higher than simple loss-based regularizers. Future research may explore more efficient ways to improve group-specific generalization, such as through specialized Network Architecture Search (NAS) or Hyperparameter Optimization (HO).

Summary updated

Abstract

Algorithmic fairness is frequently motivated in terms of a trade-off in which overall performance is decreased so as to improve performance on disadvantaged groups where the algorithm would otherwise be less accurate. Contrary to this, we find that applying existing fairness approaches to computer vision improve fairness by degrading the performance of classifiers across all groups (with increased degradation on the best performing groups). Extending the bias-variance decomposition for classification to fairness, we theoretically explain why the majority of fairness classifiers designed for low capacity models should not be used in settings involving high-capacity models, a scenario common to computer vision. We corroborate this analysis with extensive experimental support that shows that many of the fairness heuristics used in computer vision also degrade performance on the most disadvantaged groups. Building on these insights, we propose an adaptive augmentation strategy that, uniquely, of all methods tested, improves performance for the disadvantaged groups.
BibTeX
@inproceedings{zietlow2022leveling,
  author = {Dominik Zietlow and Michael Lohaus and Guha Balakrishnan and Matthäus Kleindessner and Francesco Locatello and Bernhard Schölkopf and Chris Russell},
  title = {{Leveling Down in Computer Vision: Pareto Inefficiencies in Fair Deep Classifiers}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2022},
}
Figure