An investigation into Bias in Facial Recognition using Learning Algorithms
Annye Braca · NORMA · 2017
Data Science has become a field of major activity in recent years. Whether it involves a clients propensity to churn, making medical diagnoses or inferring dating tendencies, the areas of classification and prediction are inextricably linked to this field. This research project involved the investigation of the potential to apply artificial intelligence as a discriminator of criminal tendencies in people. In addition to creating a model that attempts to classify criminals and non criminal images, the research was also concerned with the potential biases that may exist within a model of this nature. A deep learning algorithm was utilised for the classification problem and applied to numerous image subsets with consideration given to gender and emotion. Once converted to arrays, the images were high dimensional and principal components analysis was conducted as a dimension reduction technique. Results included stratified 10-fold cross validation accuracies, confusion matrices, learning curves and emotion histograms.