Impact of Explainable AI on Reduction of Algorithm Bias in Facial Recognition Technologies

Artificial Intelligence (AI) has grown dramatically over the past few decades and has much greater influence today on human performance in every walk of life. AI is great for accelerating our work, but there are also downsides to it. One such downside is the application of Facial Recognition Technologies (FRT) available today, which has adverse consequences like racial discrimination or wrongful judgement by commercial firms and even law enforcement organizations. The landmark ‘Gender Shades’ project [1] in 2018 showed a highly skewed error bar between facial recognition accuracies of subjects who are black females compared to white males using an intersectional comparison approach between facial recognition software by IBM, Microsoft and Face++. This supplemented previous studies which highlighted issues with gender classification or on the whole misidentification issues posed by facial recognition technologies (FRT) [2]. This algorithm bias towards misgendering or misclassification based on skin color could be attributed to a highly unbalanced training datasets that lack diversity as well as a lack of understanding of the black box machine learning (ML) models that are used to build the FRTs. From a Code of Ethics perspective, such algorithms do not conform to the fundamental canons [3] such as prioritizing the safety, and welfare of the public. Furthermore, they cannot be used honorably, responsibly and ethically to enhance the reputation, and usefulness of the companies promoting these technologies because they are unable to provide truthful and objective information regarding facial recognition. A possible remedy could be to curate an exhaustively diverse dataset w.r.t both color and gender. However, the task of building such a dataset would require access to a diverse population, which is not always possible in a multi-racial and multi-ethnic society. We thus propose an exhaustive review-based study of existing work on the introduction of explainable AI [4],[5] (XAI) techniques in such ML models to understand how the model itself learns. This study would explore what constraints to put on the learning method of the model, which can enforce reduction in misclassification errors from gender and skin color thus enforcing various ethical aspects like taking care of algorithm bias, introducing transparency and accountability. It would also explore and underline the overall social impact improved FRTs might have from a code of ethics perspective.