Communication-Efficient Edge AI: Algorithms and Systems (2002.09668v1)

Abstract: AI has achieved remarkable breakthroughs in a wide range of fields, ranging from speech processing, image classification to drug discovery. This is driven by the explosive growth of data, advances in machine learning (especially deep learning), and easy access to vastly powerful computing resources. Particularly, the wide scale deployment of edge devices (e.g., IoT devices) generates an unprecedented scale of data, which provides the opportunity to derive accurate models and develop various intelligent applications at the network edge. However, such enormous data cannot all be sent from end devices to the cloud for processing, due to the varying channel quality, traffic congestion and/or privacy concerns. By pushing inference and training processes of AI models to edge nodes, edge AI has emerged as a promising alternative. AI at the edge requires close cooperation among edge devices, such as smart phones and smart vehicles, and edge servers at the wireless access points and base stations, which however result in heavy communication overheads. In this paper, we present a comprehensive survey of the recent developments in various techniques for overcoming these communication challenges. Specifically, we first identify key communication challenges in edge AI systems. We then introduce communication-efficient techniques, from both algorithmic and system perspectives for training and inference tasks at the network edge. Potential future research directions are also highlighted.

Communication-Efficient Edge AI: Algorithms and Systems

The paper "Communication-Efficient Edge AI: Algorithms and Systems" by Yuanming Shi et al. provides a detailed survey on recent advancements in designing communication-efficient artificial intelligence systems deployed at the network edge. The authors examine both algorithmic and system architecture perspectives for overcoming communication challenges in edge AI tasks, emphasizing reducing the heavy data exchange overhead typically required for AI model training and inference in edge environments.

The increase in data generation by the widespread adoption of edge devices such as smartphones and tablets presents an opportunity to develop high-accuracy models locally, thus reducing the necessity for constant data transmission to a cloud infrastructure. However, this requires efficient communication frameworks to manage distributed data and model training processes over resource-constrained edge nodes. To address these challenges, the authors categorize and review techniques designed to optimize communication costs, segregating them into algorithm-level and system-level solutions.

Algorithm-Level Solutions:

1. Zeroth-Order Methods: These methods are particularly advantageous in scenarios where derivative information is unavailable or hard to compute. By estimating gradients with function values, these methods limit the data communicated back to a central node, thereby minimizing communication costs in training scenarios where only function evaluations can be performed.
2. First-Order Methods: These are widely recognized in machine learning, especially using SGD. Efforts have been made to optimize communication by reducing the number of rounds and bandwidth through strategies such as mini-batching, gradient quantization, and sparsification. By reusing or quantizing gradients, first-order methods achieve better communication efficiency, which is crucial for practical deployments over bandwidth-limited channels.
3. Second-Order Methods: By approximating rather than directly computing the Hessian matrix, second-order methods leverage curvature information to accelerate convergence, effectively reducing communication rounds albeit with higher per-round computational complexity. Variants like stochastic quasi-Newton methods demonstrate potential in trading local computation for reduced communication requirements.
4. Federated Optimization: Emphasizing privacy and model disparity preservation, federated optimization frameworks like Federated Averaging and FedProx reduce the frequency of communication by performing substantial local computation before aggregating results. However, such frameworks necessitate careful control over participant variability and data heterogeneity to maintain model performance.

System-Level Solutions:

1. Data Partition-Based Edge Training Systems: These systems distribute data workloads across various edge devices, employing both distributed and decentralized approaches. A notable advancement is the exploitation of over-the-air computation, mitigating broadcast-related delays by leveraging signal superposition properties for efficient data aggregation.
2. Model Partition-Based Edge Training Systems: Particularly effective for large models, partitioning AI models across multiple devices can significantly decrease computational and storage constraints at individual nodes. Privacy considerations are also addressed, with the vertical federated learning architecture ensuring data attribute isolation while facilitating collaborative model training.
3. Computation Offloading Based Edge Inference Systems: To overcome inherent computational limitations of edge devices, computation offloading to edge servers becomes imperative. Strategies such as partial data transmission, feature encoding, and cooperative downlink transmission improve bandwidth efficiency and reduce latency for real-time inference tasks.

In conclusion, the paper systematically dissects the interplay between algorithmic designs and system architectures in facilitating efficient AI computations at the edge. Despite notable advancements, the complexities of edge AI necessitate continued exploration of adaptive strategies aligned with emerging technologies such as 6G. The integration of edge AI into service-oriented architectures presents a frontier for scalable and sustainable intelligent systems capable of harnessing ubiquitous data responsibly and efficiently. Future work will likely delve further into hardware design, software platforms, and the realization of AI as a flexible service layer at the network edge, ensuring seamless, low-latency AI service delivery in increasingly dynamic environments.