CoCo-Fed: A Unified Framework for Memory- and Communication-Efficient Federated Learning at the Wireless Edge

TL;DR

CoCo-Fed addresses the dual bottlenecks of edge Federated Learning at wireless gNBs by (1) constraining local memory through double-dimension gradient down-projection and (2) drastically reducing backhaul traffic via orthogonal subspace superposition of layer updates. It introduces a seed-based, low-rank gradient projection coupled with a compressed-optimizer in the $r\times r$ space, and a Johnson-Lindenstrauss–style global aggregation that superimposes multiple layer updates into a single $r_a\times r_a$ matrix, recoverable per layer at the CPU. The authors provide convergence proofs for local low-rank updates under unsupervised learning, characterize matrix-combination properties with Gaussian projections, and bound the global aggregation error relative to FedAvg, showing the method converges as quantization or projection accuracy improves. Case studies on unsupervised AoA estimation demonstrate substantial memory and backhaul savings (e.g., local memory down to $O(r^2)$ and total payload to $O(r_a^2)$ per gNB) with robust convergence under IID and non-IID data, marking a practical path for native edge intelligence in O-RAN systems.

Abstract

The deployment of large-scale neural networks within the Open Radio Access Network (O-RAN) architecture is pivotal for enabling native edge intelligence. However, this paradigm faces two critical bottlenecks: the prohibitive memory footprint required for local training on resource-constrained gNBs, and the saturation of bandwidth-limited backhaul links during the global aggregation of high-dimensional model updates. To address these challenges, we propose CoCo-Fed, a novel Compression and Combination-based Federated learning framework that unifies local memory efficiency and global communication reduction. Locally, CoCo-Fed breaks the memory wall by performing a double-dimension down-projection of gradients, adapting the optimizer to operate on low-rank structures without introducing additional inference parameters/latency. Globally, we introduce a transmission protocol based on orthogonal subspace superposition, where layer-wise updates are projected and superimposed into a single consolidated matrix per gNB, drastically reducing the backhaul traffic. Beyond empirical designs, we establish a rigorous theoretical foundation, proving the convergence of CoCo-Fed even under unsupervised learning conditions suitable for wireless sensing tasks. Extensive simulations on an angle-of-arrival estimation task demonstrate that CoCo-Fed significantly outperforms state-of-the-art baselines in both memory and communication efficiency while maintaining robust convergence under non-IID settings.

Figures (7)

• Figure 1: Illustration of the architecture and operational workflow of our proposed CoCo-Fed framework.
• Figure 2: Illustration of the details of the compression and combination of the update increment of each gNB-$k$.
• Figure 3: Comparison of the convergence performance. The solid lines represent the mean MSE across multiple experimental repetitions, while the shaded areas indicate the standard deviation.
• Figure 4: Comparison between our proposed CoCo-Fed and the baselines in terms of the tradeoff between resource consumption and performance. (a) Testing MSE versus rank $r$, and (b) Testing MSE versus transmission overhead
• Figure 5: Performance comparison of the proposed CoCo-Fed and the baselines.

Theorems & Definitions (8)

• Definition 1: Reversibility Tian2020Understanding
• Theorem 1: Gradient form of reversible models
• Definition 2: $L$-continuity Zhao2024GaLore
• Theorem 2: Convergence of NN with gradient projections
• Lemma 1: JL Lemma Sanjoy2003An
• Theorem 3: Orthogonality of Random Gaussian Matrix
• Corollary 1: JL Lemma with Random Gaussian Matrix
• Theorem 4