Wednesday 26 March 2025
The EdgeCL framework, a novel approach to enabling cross-domain continual learning for edge intelligence in wireless integrated sensing and communications (ISAC) networks, has been proposed by researchers at Nanyang Technological University. The framework tackles the challenge of limited memory and computational resources on edge devices (EDs), which cannot support joint re-training on all domain datasets.
Wireless ISAC networks are becoming increasingly important for various applications such as human activity recognition, channel state information-based sensing, and more. However, these networks face significant challenges in terms of data storage and processing due to the limited resources available on EDs. To address this issue, researchers have turned to continual learning, a subfield of machine learning that enables models to learn from new data without forgetting previously learned knowledge.
The EdgeCL framework leverages transformer-based discriminators for handling sequences of noisy and non-equidistant channel state information (CSI) samples. It also employs a distilled core-set based knowledge retention method with robustness-enhanced optimization to mitigate the decline in accuracy due to catastrophic forgetting. In other words, EdgeCL allows EDs to learn from new data while preserving their ability to recall previously learned patterns.
The framework’s performance was evaluated on a real-world ISAC human activity recognition dataset, and the results showed that EdgeCL achieved 89% of the cross-domain training performance while consuming only 3% of the cumulative training memory. Moreover, the accuracy decline due to catastrophic forgetting was mitigated by 79%.
One of the key innovations in EdgeCL is its ability to learn from noisy and non-equidistant CSI samples, which are common characteristics of real-world wireless sensing data. The framework uses a transformer-based discriminator to effectively handle these imperfections and improve the overall performance.
Another significant advantage of EdgeCL is its ability to retain knowledge through a distilled core-set based method. This approach ensures that previously learned patterns are not forgotten as new data is incorporated, which is critical in applications where edge devices need to adapt to changing environments or learn from new types of data.
The EdgeCL framework has the potential to revolutionize the field of wireless ISAC networks by enabling more efficient and effective learning from real-world data. By allowing EDs to learn from new data while preserving their ability to recall previously learned patterns, EdgeCL can improve the accuracy and reliability of various applications, including human activity recognition, channel state information-based sensing, and more.
Cite this article: “Edge Intelligence in Wireless Integrated Sensing and Communications: A Novel Framework for Cross-Domain Continual Learning”, The Science Archive, 2025.
Edgecl, Continual Learning, Wireless Integrated Sensing And Communications, Edge Intelligence, Transformer-Based Discriminators, Core-Set Based Knowledge Retention, Robustness-Enhanced Optimization, Catastrophic Forgetting, Human Activity Recognition, Channel State
