Anay Majee, Rishabh Iyer
SHaSaM is a novel approach that improves fairness in facial attribute recognition by using submodular hard sample mining to address data imbalance and reduce bias from sensitive attributes.
Facial recognition systems often show bias because they learn from data that might not be balanced across different demographic groups. This can lead to unfair predictions based on attributes like race or gender. The SHaSaM method is designed to make these systems fairer by focusing on how they learn from difficult or 'hard' examples. It uses a smart selection process to balance the data and a special way of learning that minimizes the impact of sensitive attributes, making the system fairer without losing accuracy. Tests showed that it outperformed current methods, improving fairness and accuracy.
