FairALM: Augmented Lagrangian Method for Training Fair Models with Little Regret.

Area of Science:

• Computer Vision
• Machine Learning
• Algorithmic Fairness

Background:

• Algorithmic decision-making systems are increasingly prevalent.
• Concerns exist regarding biases in these models, leading to unfair treatment of certain population segments.
• Fairness-oblivious training on biased datasets results in unfair models.

Purpose of the Study:

• To explore mechanisms for incorporating fairness measures during the de novo design or training of machine learning models.
• To propose and analyze strategies for imposing fairness concurrently with model training.
• To offer a simpler alternative to existing fairness-based approaches in computer vision.

Main Methods:

• Investigated optimization concepts for imposing fairness during model training.
• Developed a routine that requires only the specification of the protected attribute.
• Compared the proposed method with adversarial training approaches.

Main Results:

• Demonstrated that fairness measures can be reliably imposed on various vision training tasks.
• Showcased an interpretable method for achieving fairness.
• Validated the effectiveness of the proposed optimization-based strategy.

Conclusions:

• The proposed method provides a simpler and interpretable way to impose fairness concurrently with model training.
• This approach effectively addresses concerns about algorithmic bias in computer vision.
• It offers a viable alternative to complex adversarial training methods for achieving fairness.