CILIATE: Towards Fairer Class-based Incremental Learning by Dataset and Training Refinement

Incremental learning systems develop fairness problems as they update: When AI models are updated with new data while retaining limited old data (a common practice to save storage and computation), they develop both dataset bias and algorithmic bias that harm fairness—a problem the paper calls "model aging."

Existing fairness solutions in incremental learning are incomplete: Current methods only partially address fairness issues in class-based incremental learning, leaving significant gaps that allow discriminatory patterns to persist as models evolve.

Strategic data selection can substantially improve fairness: CILIATE's approach of identifying and prioritizing overlooked but important samples improved fairness by 17-32% compared to leading existing methods, demonstrating that thoughtful dataset refinement during model updates is critical for responsible AI.

Fairness requires attention at multiple levels: The paper shows that fairness problems in incremental learning stem from both the data itself and how the model is trained, meaning responsible AI updates must address both sources of bias rather than treating them separately.

Practical AI systems need fairness-aware update procedures: As organizations continuously update their AI models with new data, this research suggests they should implement fairness checks and refinement processes—not just focus on accuracy—to prevent discriminatory outcomes from accumulating over time.