The widespread deployment of deep learning in real-world applications has prompted a paradigm shift toward collaborative learning between resource-constrained edge devices and powerful cloud-based infrastructures. Traditional deep learning architectures, typically optimized for centralized cloud environments, often fall short in scenarios where latency, privacy, energy efficiency, and real-time responsiveness are critical. Conversely, purely on-device models are limited in capacity and accuracy due to stringent computational and memory constraints. To address these challenges, a hybrid approach has emerged, where lightweight on-device models collaborate with large, high-capacity models hosted in the cloud, enabling the seamless integration of low-latency inference and high-accuracy computation. This survey provides a comprehensive examination of collaborative learning frameworks that bridge the gap between on-device and cloud-based models, including federated learning, split learning, model offloading, and knowledge distillation. We analyze the theoretical foundations of each approach, explore their mathematical formulations, and discuss their practical trade-offs in terms of communication overhead, privacy guarantees, learning efficiency, and robustness to heterogeneous environments.The survey further explores the optimization of collaborative inference workflows, where inference is partitioned between devices and the cloud to minimize latency and energy consumption while maximizing model accuracy. Techniques such as dynamic model partitioning, early exit strategies, quantization, and sparsification are discussed in detail, along with system-level co-design considerations that align learning objectives with hardware and network capabilities. We highlight key applications across diverse domains, including healthcare, autonomous systems, smart cities, and personalized AI, demonstrating how collaborative learning enables responsive, context-aware, and privacy-preserving AI services. Through in-depth case studies, we illustrate how these systems are implemented in practice, shedding light on architectural decisions, model deployment strategies, and real-time performance outcomes.Moreover, the abstract delves into emerging trends that are shaping the future of collaborative learning, such as privacy-enhancing technologies, edge AI hardware acceleration, continual learning, and adaptive collaboration mechanisms. The intersection of these advances is poised to redefine how AI is deployed at scale, enabling intelligent systems that are not only accurate and efficient but also secure, autonomous, and adaptable to dynamic environments. We conclude by identifying key open challenges, including the need for standardized benchmarks, scalable learning protocols, and trust frameworks that ensure responsible AI deployment in collaborative settings. This survey aims to serve as both a foundational reference for researchers entering the field and a strategic guide for practitioners designing next-generation AI systems that leverage the full potential of collaborative learning across the edge-cloud continuum.
Nadine et al. (Tue,) studied this question.