Can we trust fairness?
Although neural networks achieve incredible performance in various domains and are actively employed in high societal impact areas, their unfairness still remains a problem. Most existing computational techniques for limiting unfairness are constraint-based or regularization-based. However, often these methods work in unexpected ways and achieve fairness by introducing more complex biases into decision-making algorithm. The goal of this project is to investigate unintended consequences of fairness constraints applied to convolutional neural networks and facial recognition systems in particular.