Fairness Indicators for Systematic Assessments of Visual Feature Extractors

Priya Goyal, Adriana Romero-Soriano, Caner Hazirbas, Levent Sagun, Nicolas Usunier

Introduction

As computer vision (CV) systems transition from controlled research environments to large-scale social deployment, the risk of algorithmic bias has become a critical concern. Biases often manifest as misidentification of individuals or degraded performance across specific demographic groups. Traditional fairness audits typically treat models as "black boxes," evaluating the final output of a specific application, such as a facial recognition system. However, modern CV architectures are built upon foundational feature extractors—backbones pre-trained on massive datasets to generate mathematical embeddings.

If these foundational models harbor biases, every downstream application built upon them inherits those flaws. To address this, researchers from Meta AI propose a standardized framework of "fairness indicators" designed to audit the "fairness health" of feature extractors before deployment.

Methodology: The Fairness Indicator Framework

The authors introduce three distinct indicators to probe various dimensions of algorithmic harm. This systematic approach allows researchers to evaluate models without necessarily training exhaustive downstream classifiers for every possible use case.

Figure 1: Conceptual overview of the assessment pipeline, from diverse datasets through a pre-trained feature extractor to specific fairness indicators.

The framework relies on a suite of diverse datasets, including Casual Conversations (CC) for self-identified demographics, OpenImages MIAP for perceived attributes, UTK Faces for representation analysis, and Dollar Street for cross-cultural object recognition.

The Three Indicators

The proposed indicators are summarized in the table below, detailing their specific goals and the metrics used for evaluation.

Figure 3: Summary of the three proposed fairness indicators, including their target tasks, goals, sensitive groups, and evaluation metrics.
  1. Harmful Label Association: This indicator measures the propensity of a model to assign offensive or stereotypical labels (e.g., "animal" or "crime") to images of people. By mapping ImageNet classes into a hierarchy (Human vs. Non-Human), the authors quantify mislabeling rates across demographics.
  2. Geographical Diversity: This assesses the model's ability to recognize everyday objects (e.g., soap, stoves) across different regions and household income levels. It highlights whether models trained on Western-centric data fail in lower-income or non-Western contexts.
  3. Same-group Attribute Assessment: This is a direct probe of the feature space using similarity search. By identifying the "nearest neighbors" of an image embedding, the researchers check if the model naturally clusters people by sensitive attributes like gender or skin tone, revealing biases baked into the underlying geometry of the representation.

Key Insights and Results

The study compared traditional supervised learning on curated datasets (like ImageNet) against Self-Supervised Learning (SSL) and Weakly-Supervised Learning (WSL) on massive, uncurated datasets (e.g., Instagram).

The Superiority of Self-Supervised Learning

A primary finding is that SSL models, such as SEER, often yield fairer representations than supervised models. Despite being trained on uncurated data, SSL models showed fewer harmful associations and superior geographical diversity. This suggests that the objective of SSL—learning general features without explicit labels—may inherently mitigate some forms of stereotypical bias found in human-annotated datasets.

Domain and Content over Scale

While increasing data scale (from 1M to 3.6B images) generally improves performance, the content of the training data is more influential. Models trained on human-centric data (Instagram) performed significantly better on fairness tasks involving people compared to those trained on the object-centric ImageNet.

The Confidence and Income Gaps

The research identified a "confidence gap": supervised models are more likely to "hallucinate" stereotypical labels with high certainty, whereas SSL models tend to be more cautious. Furthermore, a blatant "income gap" exists in object recognition; models were approximately 25% more accurate on items from high-income households than low-income ones. However, SSL and WSL models were more effective at closing this disparity than supervised baselines.

Conclusion and Future Directions

This work provides a reproducible protocol for auditing the fairness of foundational CV models. However, limitations remain. The current indicators are constrained by existing dataset labels, which often rely on binary gender or perceived attributes. Additionally, the scope is currently limited to single-label classification and similarity search.

Future research should extend these indicators to multi-label tasks and object detection. By releasing the code for these indicators, the authors aim to foster a community standard for fairness auditing, ensuring that the building blocks of AI are as equitable as the applications they support.

Summary updated

BibTeX
@inproceedings{goyal2022fairness,
  author = {Priya Goyal and Adriana Romero-Soriano and Caner Hazirbas and Levent Sagun and Nicolas Usunier},
  title = {{Fairness Indicators for Systematic Assessments of Visual Feature Extractors}},
  booktitle = {FAccT},
  year = {2022},
  url = {https://openreview.net/forum?id=JDFel59Yql-},
}
Figure