The Impact of Biases in Facial Recognition Artificial Neural Networks


Following a controlled experiment regarding the testing of a convolutional neural network (CNN) on the task of recognizing and classifying faces of transgender people and non-white people, preliminary data analysis has suggested the need to further incorporate transgender people into datasets when training facial recognition neural networks. The CNN model used in this experiment is a pre-trained model, which was thus tested on 3 different datasets in order to measure potential biases: a novel dataset consisting of self-reported binary transgender individuals, a balanced dataset, and an unbalanced dataset. Similar to research suggested by prominent authors in the field of AI - specifically regarding the potential dangers of biases in such algorithms - it was found that self-identifying binary transgender men were more often misgendered than self-identifying binary transgender women. Further research is needed in order to potentially mitigate such biases in future iterations of neural networks.

Apr 19, 2023 8:00 AM — 6:00 PM

For further information on the topic of this project, please refer to the AI Bias project page.