A Review of Continual Learning in Edge AI

The transition from laboratory-controlled AI training to real-world deployment reveals a critical gap: traditional episodic training paradigms fail to address the dynamic, resource-constrained nature of edge environments. Sustainable continual intelligence represents a transformative paradigm that integrates adaptive learning capabilities with sustainability principles for edge AI systems, enabling perpetual learning and evolution within resource constraints while maintaining operational effectiveness. This paper provides a review of continual learning in edge AI, which examines how federated architectures enable distributed systems to learn sustainably under resource constraints through network-aware coordination. Our key findings reveal that current approaches treat networking as a bottleneck rather than an enabler, missing opportunities for communication-learning co-design. This paper first emphasizes the fundamental limitations of episodic paradigms and the compelling necessity for continual learning in edge deployments. Subsequently, we examine the theoretical foundations and system design principles for always-on learning, covering continual federated learning architectures and adaptive model evolution strategies. Next, we explore implementation mechanisms and sustainability assurance across energy-aware lifecycle management and fault-tolerant operations. We systematically review practical applications, identify critical research gaps in theoretical foundations, standardization needs, and opportunities for cross-layer co-design. This work presents the first comprehensive framework for designing AI systems that can sustainably operate and improve in production edge environments, thereby bridging the gap between academic research and industrial deployment needs.