Läs rummet : Undersökning av partiskhet i maskinlärning för igenkänning av ansiktsuttryck

This study aimed to examine what biases would arise when training binary and multiclass machine learning models for facial emotion recognition with eight emotions, sad, fear, anger, surprise, happy, neutral, disgust, and contempt, on clearly biased datasets with regard to ethnicity and sex. The different models were trained on all demographics, solely white subjects, solely non-white subjects, solely male subjects, and solely female subjects. The point was to examine which machine learning algorithm between convolutional neural networks and support vector machines would suffer the most from having been trained on a biased dataset and how it could contribute to the UN’s sustainability goal 10: reduced inequalities. It was found that both algorithms received lower accuracies from training on a biased dataset, however, support vector machines had more difficulties in applying training from a biased dataset to a non-biased dataset. This result shows that to contribute to inclusivity within machine learning and therefore reach the UN’s sustainability goal 10, it is important to use a well-balanced dataset.