Representation Learning with Statistical Independence to Mitigate Bias.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Bias in datasets and tasks is a critical challenge in machine learning.
• Controlling bias during dataset curation is often difficult or impossible.
• Fair representation learning offers an alternative to address bias in existing data.

Purpose of the Study:

• To propose a novel model for fair representation learning using adversarial training.
• To develop a model that learns features with maximum task discriminative power and minimal statistical dependence on protected variables.
• To introduce a new adversarial loss function that minimizes correlation between bias and learned features.

Main Methods:

• Adversarial training with two competing objectives.
• Incorporation of a novel adversarial loss function.
• Application to synthetic data, medical images, and a gender classification dataset.

Main Results:

• The proposed method learns features that are both highly discriminative for the task and unbiased.
• Demonstrated superior prediction performance compared to existing methods.
• Successfully mitigated bias in both task-biased and dataset-biased scenarios.

Conclusions:

• The adversarial learning approach effectively addresses bias in machine learning.
• Learned features exhibit enhanced prediction accuracy and fairness.
• The method provides a viable solution for developing unbiased machine learning models.