Adaptive Deep Learning Architectures for Real-Time Data Streams in Edge Computing Environments
Abstract
The increasing reliance on real-time analytics within edge computing environments has underscored the need for adaptive deep learning architectures capable of handling continuous data streams under limited computational and energy resources. Unlike traditional cloud-based frameworks, edge computing shifts computation closer to the data source, minimizing latency, preserving data privacy, and supporting responsive decision-making in dynamic contexts. This paper comprehensively examines adaptive deep learning models that autonomously adjust their structure and parameters in response to evolving data distributions and resource constraints. The study explores several key methodologies—such as master–surrogate deep neural networks, context–adaptive DNN atom partitioning, distributed inference with fused layer partitioning, and reinforcement learning–driven resource scheduling. Comparative analyses demonstrate that these adaptive frameworks achieve substantial performance gains, including up to 23.31% accuracy improvement, 62.14% latency reduction, and over 50% energy savings across diverse edge devices. Furthermore, the paper evaluates real-world deployments in smart cities, healthcare monitoring, Industry 4.0 automation, autonomous vehicles, and environmental sensing, illustrating the scalability and robustness of adaptive architectures in heterogeneous edge ecosystems. The discussion concludes by outlining current challenges—such as hardware diversity, continual learning, and energy constraints—and highlights future research directions including federated adaptation, neuromorphic hardware integration, and standardized benchmarking for edge AI.