BiasFairnessAccountabilitycs.CV
Mitigating Algorithmic Bias on Facial Expression Recognition
Glauco Amigo, Pablo Rivas Perea, Robert J. Marks
December 23, 2023
arXiv.org
3 citations
Key Takeaways
-
Imbalanced training data leads to algorithmic discrimination: Machine learning systems trained on datasets where some categories are underrepresented will systematically perform worse for those groups, meaning facial expression recognition systems may misclassify emotions for people from minority groups more frequently than others.
-
Bias in AI disproportionately affects minority populations: The paper specifically highlights that dataset bias is "especially sensitive when dealing with minority people groups," indicating that algorithmic fairness issues have real consequences for underrepresented communities in facial recognition applications.
-
Debiasing techniques can help level the playing field: The researchers demonstrate that a "debiasing variational autoencoder" approach can help generate algorithms that "treat every person equally" despite starting with biased training data, suggesting technical solutions exist to mitigate—though not eliminate—these fairness problems.
-
The bias problem is widespread, not isolated: The paper notes that "biased datasets are ubiquitous," indicating this isn't a rare edge case but a systemic challenge across machine learning applications that developers need to actively address.
Abstract
Biased datasets are ubiquitous and present a challenge for machine learning. For a number of categories on a dataset that are equally important but some are sparse and others are common, the learning algorithms will favor the ones with more presence. The problem of biased datasets is especially sensitive when dealing with minority people groups. How can we, from biased data, generate algorithms that treat every person equally? This work explores one way to mitigate bias using a debiasing variational autoencoder with experiments on facial expression recognition.