Research
In 2023, having engaged in discussion panels about gender inequality in STEM at the Ross math program and started exploring the role of technology in perpetuating these biases, I became increasingly aware of the pervasive gender bias in technology and its impact on opportunities for women in STEM. This realization motivated me to research how machine learning could be used to mitigate some gender bias, which ultimately led to this research.
Using Deep Learning to Remove Potential Gender Bias in Computer Vision Tasks While Preserving Test Data Accuracy
Abstract: This research addresses gender bias in computer vision by evaluating various mitigation techniques and introducing a novel benchmark for analysis. Initially, limitations in popular adversarial training approaches were identified, leading to the proposal of the Reducing Bias Amplification (RBA) method as an alternative. Four models—baseline, strategic sampling, domain discriminative training, and domain-independent training—were compared, with domain-independent training emerging as the most effective. Using the CelebA dataset for validation, results confirmed that domain-independent training significantly mitigates gender bias in visual recognition, validated by multiple evaluative indicators.