Bias Amplification Improves Minority Group Performance in Machine Learning Models

Bias amplification through trainable auxiliary variables can enhance the performance of minority groups in machine learning models, even without access to group annotations during training.