Continual learning, degradation detection, and adaptation under distribution shift in optical networks

Develop continual (lifelong) learning methods for AI/ML models in optical network automation that can detect AI degradation and adapt models to progressive distribution shifts encountered in real deployments.

Background

Beyond data scarcity, the paper highlights the absence of mechanisms for models to learn continuously and remain performant as network conditions evolve. This includes detecting when model performance degrades and adapting to gradual shifts in data distributions.

Such capabilities are essential for sustainable AI operations in production optical networks and align with the paper’s emphasis on active inference as a pathway to lifelong learning.

References

Further, according to , key AI research challenges remain open: ($i$) Training: Lack of available training datasets from real-world network deployments, ($ii$) Learning: Lack of lifelong (i.e., continual) learning, including AI degradation detection and model adaptation to progressive distribution shift, and ($iii$) Explainability: Lack of trustworthy explainable AI (XAI) due to insufficient transparency of blackbox AI.

— From Artificial Intelligence to Active Inference: The Key to True AI and 6G World Brain [Invited] (2505.10569 - Maier, 29 Apr 2025) in Section 1 (Introduction), Point Alpha — Active Inference in Optical Networks