Less can be more: representational vs. stereotypical gender bias in facial expression recognition
[FULL TEXT READ at the publisher, peer-reviewed, open access CC BY. Author order verified at api.openalex.org: IRIS DOMINGUEZ-CATENA FIRST AND CORRESPONDING, Public University of Navarre; Galar last] Dominguez-Catena, I., Paternain, D., Jurio, A., & Galar, M. (2024). 'Less can be more: representational vs. stereotypical gender bias in facial expression recognition.' Progress in Artificial Intelligence. doi:10.1007/s13748-024-00345-w. THE MOST USEFUL BIAS FINDING IN THE FILE AND IT REVERSES THE USUAL STORY. Method: AffectNet as the baseline, with training subsets engineered to carry varying strengths of REPRESENTATIONAL bias, a group under-represented across all classes, and STEREOTYPICAL bias, a group under-represented in one specific class, with per-class recall computed separately by gender group on the full AffectNet test set. THEIR RESULT, VERBATIM: 'Our results show that representational bias has a weaker impact than expected. Models exhibit a good generalization ability even in the absence of one gender in the training dataset. Conversely, stereotypical bias has a significantly stronger impact, primarily concentrated on the biased class, although it can also influence predictions for unbiased classes.' The magnitude: 'In the overrepresented groups there is a weak increase in recall (under 0.1). Conversely, the impact is more pronounced in the underrepresented groups, with a total drop of up to 0.4 in the extreme cases compared to the balanced reference.' THEIR EXPLANATION, VERBATIM, and it is the disquieting part: 'In representational bias, where a demographic group is underrepresented across all classes, the model can often generalize from the overrepresented group to the underrepresented one, assuming similar emotional expressions between the groups. However, stereotypical bias induces a new pattern that can be learned by the model: the observable
[FULL TEXT READ at the publisher, peer-reviewed, open access CC BY. Author order verified at api.openalex.org: IRIS DOMINGUEZ-CATENA FIRST AND CORRESPONDING, Public University of Navarre; Galar last] Dominguez-Catena, I., Paternain, D., Jurio, A., & Galar, M. (2024). 'Less can be more: representational vs. stereotypical gender bias in facial expression recognition.' Progress in Artificial Intelligence. doi:10.1007/s13748-024-00345-w. THE MOST USEFUL BIAS FINDING IN THE FILE AND IT REVERSES THE USUAL STORY. Method: AffectNet as the baseline, with training subsets engineered to carry varying strengths of REPRESENTATIONAL bias, a group under-represented across all classes, and STEREOTYPICAL bias, a group under-represented in one specific class, with per-class recall computed separately by gender group on the full AffectNet test set. THEIR RESULT, VERBATIM: 'Our results show that representational bias has a weaker impact than expected. Models exhibit a good generalization ability even in the absence of one gender in the training dataset. Conversely, stereotypical bias has a significantly stronger impact, primarily concentrated on the biased class, although it can also influence predictions for unbiased classes.' The magnitude: 'In the overrepresented groups there is a weak increase in recall (under 0.1). Conversely, the impact is more pronounced in the underrepresented groups, with a total drop of up to 0.4 in the extreme cases compared to the balanced reference.' THEIR EXPLANATION, VERBATIM, and it is the disquieting part: 'In representational bias, where a demographic group is underrepresented across all classes, the model can often generalize from the overrepresented group to the underrepresented one, assuming similar emotional expressions between the groups. However, stereotypical bias induces a new pattern that can be learned by the model: the observable
Written from the fULL TEXT READ at the publisher, peer-reviewed, open access CC BY. Author order verified at api.openalex.org: IRIS DOMINGUEZ-CATENA FIRST AND CORRESPONDING, Public University of Navarre; Galar last.
What our sources record
Recorded when this source was checked, figures included, with page references and quotation marks removed.
Progress in Artificial Intelligence. doi:10.1007/s13748-024-00345-w