Equitable Multi-Task Learning for AI-RANs

A team of researchers including Panayiotis Raptis, Fatih Aslan, and George Iosifidis has developed a novel framework called OWO-FMTL (Online-Within-Online Fair Multi-Task Learning) specifically designed for AI-RANs—AI-enabled Radio Access Networks that will form the backbone of 6G systems. These networks must serve diverse users with varying learning tasks while sharing limited edge computing resources. The core challenge addressed is ensuring equitable inference performance across all users, preventing scenarios where some users receive significantly better AI service than others due to resource constraints or task complexity.

The OWO-FMTL framework operates through two interconnected learning loops: an outer loop that updates the shared AI model across training rounds, and an inner loop that dynamically rebalances user priorities within each round using efficient primal-dual updates. This dual-loop approach mathematically guarantees diminishing performance disparity over time, quantified through generalized alpha-fairness metrics that allow network operators to tune the trade-off between overall system efficiency and individual user fairness.

Experimental validation on both convex optimization tasks and deep learning scenarios confirms that OWO-FMTL outperforms existing multi-task learning baselines, particularly under dynamic conditions where user demands and network resources fluctuate. The framework's computational overhead remains low enough for practical edge deployment, making it suitable for real-time implementation in next-generation wireless infrastructure where milliseconds matter.