Abstract
Eye gaze is the observation and collection of images that show eye behavior, such as pupil dilation and movement, that help identifies a person’s fixation on specific points concerning screen-based media. Eye gaze could be an essential tool that could be used across many industries, yet no technology exists on a large scale. One possible way to develop this technology is by using a supervised deep-learning model such as convolutional neural networks. I used a preexisting eye gaze dataset called GazeCapture to obtain predictions of the gaze locations along with obtaining algorithm fairness metrics for the race and gender variables that were added to the dataset using the FairFace demographic classification algorithm. It is important to discover the algorithm fairness of deep learning models because many deep learning models have been unfair and harmful to underrepresented groups. To test these metrics, I used image processing methods including RGB, grayscale, binarization, and edge detection to analyze how image processing methods have an effect on an algorithm’s fairness. I found that RGB image processing is the most accurate model for predicting the gaze locations with a prediction error of 2.7172 cm for mobile devices. Then when investigating algorithm fairness, I found that the Canny edge detection resulted in a fair gender algorithm that obtained a p-value of 0.94 when determining if the error prediction for males and females were different, while only decreasing the accuracy by 0.3292 cm. Additionally, the intensity grayscale method also resulted in a fair gender algorithm that obtained a p-value of 0.09. Given a p-value threshold v of 0.05, these results are significant and the method only decreases the accuracy by 0.0568 cm. Furthermore, no image processing method resulted in a fair race algorithm. Future work should include creating a dataset that includes race and gender demographics and has an equal number of participants and images for all racial and gender groups. The code and data files are available on my Github.
Advisor
Horr, Christina
Department
Statistical and Data Sciences
Recommended Citation
Meyer, Adam P., "Uncovering Unintended Consequences: A Deep Learning Approach to Algorithmic Fairness in Eye Gaze Prediction" (2023). Senior Independent Study Theses. Paper 10684.
https://openworks.wooster.edu/independentstudy/10684
Disciplines
Data Science
Keywords
Deep Learning, Algorithm Fairness, Eye Gaze
Publication Date
2023
Degree Granted
Bachelor of Arts
Document Type
Senior Independent Study Thesis
External Link
https://github.com/ameyer23-m/Algorithm-Fairness-and-a-Deep-Learning-Approach-to-Eye-Gaze
© Copyright 2023 Adam P. Meyer