[Paper Review] Wireless for Machine Learning

Henrik Hellström, José Mairton B. da Silva+5 more|arXiv (Cornell University)|Aug 31, 2020

Distributed Sensor Networks and Detection Algorithms121 references8 citations

TL;DR

This survey proposes a comprehensive review of wireless communication techniques specifically designed to support machine learning (ML) over distributed datasets, focusing on two key approaches: analog over-the-air computation for high spectral efficiency and digital radio resource management (RRM) optimized for distributed ML. It demonstrates that these methods can significantly improve energy efficiency, latency, and spectrum utilization in ML workloads, especially in dynamic environments like UAVs and smart cities.

ABSTRACT

As data generation increasingly takes place on devices without a wired connection, machine learning (ML) related traffic will be ubiquitous in wireless networks. Many studies have shown that traditional wireless protocols are highly inefficient or unsustainable to support ML, which creates the need for new wireless communication methods. In this survey, we give an exhaustive review of the state-of-the-art wireless methods that are specifically designed to support ML services over distributed datasets. Currently, there are two clear themes within the literature, analog over-the-air computation and digital radio resource management optimized for ML. This survey gives a comprehensive introduction to these methods, reviews the most important works, highlights open problems, and discusses application scenarios.

Motivation & Objective

• Address the inefficiency of traditional wireless protocols in handling machine learning workloads due to high latency, energy consumption, and poor scalability with distributed data.
• Identify the need for new wireless communication methods tailored to machine learning, especially for federated learning and over-the-air computation.
• Review and categorize state-of-the-art techniques in analog over-the-air computation and digital radio resource management (RRM) for distributed machine learning (DML).
• Highlight open research problems in data importance modeling, CSI feedback, and integration into existing wireless infrastructure.
• Demonstrate practical applications in smart cities, UAVs, vehicular networks, virtual reality, and edge caching to illustrate real-world impact.

Proposed method

• Proposes analog over-the-air computation as a method to compute functions (e.g., averages) directly over the wireless channel, reducing latency and communication overhead by avoiding digital quantization.
• Introduces digital radio resource management (RRM) strategies tailored for federated learning, including power allocation, user scheduling, and beamforming to optimize model convergence and reduce training time.
• Analyzes CSI-free over-the-air computation methods (e.g., blind methods) to handle rapidly changing channels in mobile environments like UAV deployments.
• Introduces the use of federated learning (FL) for training dynamic models—such as optimal UAV base station positioning—using distributed channel state information from mobile devices.
• Proposes joint optimization of resource allocation and fairness (e.g., JAFMA, CRM) to balance performance and fairness in multi-user ML training scenarios.
• Applies machine learning-based channel prediction models to enable adaptive routing and coverage optimization in ad-hoc UAV networks during disaster response.

Experimental results

Research questions

• RQ1 How can over-the-air computation reduce communication overhead and latency in distributed machine learning compared to traditional digital transmission?
• RQ2 What are the key challenges in integrating analog over-the-air computation into existing 5G/6G wireless infrastructure?
• RQ3 How can radio resource management be optimized to reflect data importance and model convergence dynamics in federated learning?
• RQ4 What role can federated learning play in enabling autonomous UAV base stations to adapt their position based on real-time channel conditions?
• RQ5 How can ML-based channel prediction improve routing and connectivity in dynamic, infrastructure-less environments such as disaster zones?

Key findings

• Analog over-the-air computation achieves a theoretical throughput gain proportional to the number of users, significantly improving spectral efficiency in distributed learning.
• Digital RRM techniques optimized for FL reduce training latency and energy consumption by prioritizing high-importance data and minimizing redundant transmissions.
• CSI-free over-the-air computation enables low-latency inference in rapidly fading environments, such as those with high-mobility UAVs.
• Federated learning enables dynamic UAV base station positioning by training on distributed channel state information, improving coverage and data rates in real time.
• ML-based channel prediction models can enhance routing resilience in multi-hop UAV ad-hoc networks, especially in post-disaster scenarios with no existing infrastructure.
• The integration of wireless-ML techniques in smart cities, VR, and edge caching can reduce backhaul load and improve QoS, with measurable gains in energy and spectrum efficiency.