Despite the numerous promising advancements in the field of artificial intelligence, algorithmic bias is pervasive in modern deep learning systems, and it is often ignored. Biometric AIs have been identified to contain racial bias, gender bias, age bias, and more. This thesis is a study of the bias that exists in face-based soft biometric systems. Three solutions are proposed for the purpose of mitigating bias. The first solution involves data augmentation as a method of mitigating bias. Trained models are usually biased against dataset minorities, and since data augmentation is essentially a method of manufacturing more training data, it is proposed that data augmentation will help glean more information from these minorities. The second solution involves oversampling parts of the dataset. This gives a model a more even representation across different classes so that error is not simply optimized for the majority. The third solution involves the use of generated data to synthetically help balance the training set. Each of these solutions are explored for their efficacy of bias mitigation. State-of-the-art results are obtained for age, gender, and BMI estimation.

Recommended Citation: Smith P., Ricanek K., Kline D., Mittal S., (2020). Mitigating Algorithmic Bias in Deep Convolutional Neural Networks. UNCW MS CSIS Proceedings. V. 14 , N. 5 .