Fair Class-Incremental Learning using Sample Weighting

Fairness gaps widen as AI systems learn new tasks: The paper identifies that when machine learning models learn new information sequentially (class-incremental learning), they tend to "forget" how to perform well for certain sensitive groups—like specific demographic categories or underrepresented classes—more than others, creating fairness problems that standard accuracy-focused approaches miss.

Mathematical root cause identified: Researchers discovered that unfair forgetting happens when the direction of learning updates for new tasks conflicts with the learning direction needed to maintain performance for sensitive groups, which they measure through mathematical inner products of gradient vectors.

Practical solution through sample weighting: The proposed Fairness-aware Sample Weighting (FSW) algorithm adjusts how much importance the model gives to different training examples, essentially telling the system to pay more attention to data from groups that would otherwise be forgotten, without requiring major changes to existing systems.

Better balance between accuracy and fairness achieved: Experiments demonstrate that FSW improves the accuracy-fairness tradeoff compared to existing methods, meaning organizations can now maintain overall model performance while ensuring underrepresented groups aren't left behind as systems learn new tasks.

Scalable approach using established optimization methods: The framework can be solved using linear programming, a well-established mathematical technique, making it computationally efficient and practical for real-world deployment in trustworthy AI systems.