Analysis of racial bias in facial emotion recognition

The complexity of deep learning models with high-dimensional inputs and subjective labels extends to fairness. In facial emotion recognition, where datasets can be racially unbalanced, models may produce biased results for different racial groups. This raises concerns about fairness and highlights the need for further research on racial bias. Inaccurate generalization owing to such bias could negatively affect real-world performance. This thesis aims to evaluate racial bias using subsampling techniques to create training sets with diverse racial distributions and then examine test performances across these experiments using state-of-the-art facial emotion recognition techniques.