analiza aiqbench
This paper provides a formal theoretical framework explaining how standard gradient descent in deep learning inherently amplifies bias toward majority groups when population or variance imbalances exist. The authors demonstrate that training on full datasets often results in stereotypical predictors that neglect minority-specific features, establishing a lower bound on the additional training required to mitigate this effect. These findings offer critical insights into the optimization dynamics of machine learning models and the structural causes of representation bias.
When majority rules, minority loses: bias amplification of gradient descent
arXiv:2505.13122v3 Announce Type: replace-cross Abstract: Despite growing empirical evidence of bias amplification in machine learning, its theoretical foundations remain poorly understood. We develop a formal framework for majority-minority learning tasks, showing how standard training can favor majority groups and produce stereotypical predictors that neglect minority-specific features. Assuming population and variance imbalance, our analysis reveals three key findings: (i) the close proximity between ``full-data'' and stereotypical predictors, (ii) the dominance of a region where training the entire model tends to merely learn the majority traits, and (iii) a lower bound on the additional training required. Our results are illustrated through experiments in deep learning for tabular and image classification tasks.
Abstract:Despite growing empirical evidence of bias amplification in machine learning, its theoretical foundations remain poorly understood. We develop a formal framework for majority-minority learning tasks, showing how standard training can favor majority groups and produce stereotypical predictors that neglect minority-specific features. Assuming population and variance imbalance, our analysis reveals three key findings: (i) the close proximity between ``full-data'' and stereotypical predictors, (ii) the dominance of a region where training the entire model tends to merely learn the majority traits, and (iii) a lower bound on the additional training required. Our results are illustrated through experiments in deep learning for tabular and image classification tasks.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Optimization and Control (math.OC)
arXiv:2505.13122 [cs.LG]
(or arXiv:2505.13122v3 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2505.13122 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 203 cuvinte
