Reproduce “Auditing Facial Emotion Recognition Datasets for Posed Expressions and Racial Bias” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Rina Khan, Catherine Stinson

Researchers audited common "in-the-wild" facial emotion recognition datasets for the prevalence of posed expressions and analyzed state-of-the-art models for racial bias. The study identified a high proportion of posed expressions in the datasets and demonstrated that models frequently misclassify neutral or positive expressions of non-White and darker-skinned individuals as negative emotions.

Key insights

  • Current 'in-the-wild' FER datasets contain a substantial percentage of posed expressions (over 40%), challenging their ecological validity and the true performance of models trained on them.
  • State-of-the-art FER models exhibit racial bias, consistently misclassifying smiling or neutral expressions of non-White and darker-skinned individuals as negative emotions more often than for White individuals.
  • These algorithmic biases align with existing societal stereotypes and have the potential to cause harm in high-stakes real-world applications, emphasizing the need for a re-evaluation of FER's theoretical underpinnings and ethical development.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.