Abstract

The Cloud-Edge Continuum (CEC) represents a paradigm shift towards a heterogeneous, distributed computing landscape. This environment is characterized by massively dis-tributed data sources, dynamic network conditions, and fluctuating computational loads. Traditional Machine Learning (ML) models, trained offline in a centralized manner, are not suited for this reality. They fail to adapt to the constant stream of new data, making them vulnerable to concept drift. This leads to inevitable performance degradation, creates significant processing bottlenecks, and undermines core Trustworthy AI principles of robustness and reliability. This paper argues that Continual Learning (CL) is a critical and necessary paradigm for robust and efficient intelligence in the CEC. We review the relevance of CL, data stream learning, and integrated concept drift detection as the primary mechanisms for maintaining model robustness and resilience. CL, implemented through a combination of data-centric and model-centric compression and frugal AI techniques, is vital for achieving both the efficiency and trustworthiness demanded by next-generation applications operating in the CEC. These methodologies include iterative fine-tuning, model compression, knowledge distillation, and dynamic neural network growth. This adaptive intelligence is required not only for end-user applications but also for MetaOS-level orchestration across the Cloud–Edge Continuum. This paper concludes by presenting key findings that highlight the essential role of adaptive learning across the continuum and outlines future research directions aimed at enabling scalable, trustworthy, and resource-efficient continual learning for MetaOS-based orchestration and management.

Keywords:Continual learning; Cloud-Edge continuum; Concept drift; Trustworthy AI; Frugal AI; Data stream learning; Distillation; Edge AI

Abbreviations: CEC: Cloud-Edge Continuum; ML: Machine Learning; P2P: Peer-to-Peer; RAG: Retrieval- Augmented Generation; ICL: In-Context Learning; PEFT: Parameter-Efficient Fine-Tuning; MoE: Mixture of Experts; GNG: Growing Neural Gas; FL: Federated Learning; DRL: Distributed Reinforcement Learning; MetaOS: Meta-Operating System