arXiv subjects

Jaeyoung Song

Publications and source records attributed to Jaeyoung Song.

8 recordsLinked to original sources

arXiv2026

Bandwidth Allocation with Device Partitioning for Federated Learning over Industrial IoT networks

We consider a federated learning (FL) system in which Industrial Internet-of-Things (IIoT) devices collaboratively train a global model over wireless channels without sharing local data. In such systems, communication time is a primary bottleneck that constrains overall training efficiency. Unlike conventional networks that prioritize individual quality-of-service requirements, FL systems collectively aim to converge to an optimal global model as efficiently as possible, which calls for a fundamentally different approach to bandwidth allocation. In this paper, we propose a novel bandwidth allocation policy that exploits the heterogeneity of device computing capabilities to minimize total training time. Rather than distributing bandwidth among all selected devices simultaneously, the proposed policy partitions the participating devices into ordered subsets and sequentially grants each subset exclusive access to the full bandwidth. We formally prove that this partitioning-based policy achieves a strictly lower training time than any bandwidth allocation scheme without partitioning, irrespective of the underlying scheduling algorithm. Furthermore, by reducing per-device transmission duration, the proposed policy also minimizes uplink energy consumption, which is particularly beneficial for battery-constrained IIoT devices. Extensive experiments on real-world datasets - including GC10-Det, an industrial surface defect benchmark, and CIFAR-10, a standard image classification benchmark - demonstrate that the proposed policy consistently reduces training time and energy consumption compared to existing bandwidth allocation schemes, approaching the theoretical lower bound on round time.

cs.LG↗

arXiv2026

Separate Aggregation of Split Network for Personalized Federated Learning

Federated learning enables collaborative model training without sharing raw data, but its performance can degrade substantially under heterogeneous client data distributions. A single global model often cannot satisfy diverse client requirements, so personalized federated learning has therefore been explored to improve client specific performance while preserving global generalization. Existing PFL methods often face a fundamental tradeoff in which stronger global sharing can undermine local specialization, whereas stronger local adaptation can lead to overfitting under limited data, label imbalance, and missing class scenarios. In this work, we propose PGFedSplit, a personalized federated learning framework that improves both personalization and global generalization under severe client heterogeneity. PGFedSplit adopts a split architecture and performs adaptive aggregation scheduling tailored to the roles of different model components, enabling stable knowledge sharing while maintaining client specific adaptation. Each client further leverages a mixture of locally extracted representations and synthetic representations generated from server side Gaussian statistics, improving robustness under label imbalance and missing class conditions. Extensive experiments on Fashion MNIST, CIFAR 10, CIFAR 100, and Tiny ImageNet demonstrate consistent improvements over state of the art PFL methods, with stable convergence and superior personalization in highly heterogeneous settings.

cs.LG↗

arXiv2024

Asynchronous Federated Learning Using Outdated Local Updates Over TDMA Channel

In this paper, we consider asynchronous federated learning (FL) over time-division multiple access (TDMA)-based communication networks. Considering TDMA for transmitting local updates can introduce significant delays to conventional synchronous FL, where all devices start local training from a common global model. In the proposed asynchronous FL approach, we partition devices into multiple TDMA groups, enabling simultaneous local computation and communication across different groups. This enhances time efficiency at the expense of staleness of local updates. We derive the relationship between the staleness of local updates and the size of the TDMA group in a training round. Moreover, our convergence analysis shows that although outdated local updates hinder appropriate global model updates, asynchronous FL over the TDMA channel converges even in the presence of data heterogeneity. Notably, the analysis identifies the impact of outdated local updates on convergence rate. Based on observations from our convergence rate, we refine asynchronous FL strategy by introducing an intentional delay in local training. This refinement accelerates the convergence by reducing the staleness of local updates. Our extensive simulation results demonstrate that asynchronous FL with the intentional delay can rapidly reduce global loss by lowering the staleness of local updates in resource-limited wireless communication networks.

cs.DC↗

arXiv2024

Optimal Batch Allocation for Wireless Federated Learning

Federated learning aims to construct a global model that fits the dataset distributed across local devices without direct access to private data, leveraging communication between a server and the local devices. In the context of a practical communication scheme, we study the completion time required to achieve a target performance. Specifically, we analyze the number of iterations required for federated learning to reach a specific optimality gap from a minimum global loss. Subsequently, we characterize the time required for each iteration under two fundamental multiple access schemes: time-division multiple access (TDMA) and random access (RA). We propose a step-wise batch allocation, demonstrated to be optimal for TDMA-based federated learning systems. Additionally, we show that the non-zero batch gap between devices provided by the proposed step-wise batch allocation significantly reduces the completion time for RA-based learning systems. Numerical evaluations validate these analytical results through real-data experiments, highlighting the remarkable potential for substantial completion time reduction.

cs.LG↗

arXiv2023

Optimal Scheduling Policy for Minimizing Age of Information with a Relay

We consider IoT sensor network where multiple sensors are connected to corresponding destination nodes via a relay. Thus, the relay schedules sensors to sample and destination nodes to update. The relay can select multiple sensors and destination nodes in each time. In order to minimize average weighted sum AoI, joint optimization of sampling and updating policy of the relay is investigated. For errorless and symmetric case where weights are equally given, necessary and sufficient conditions for optimality is found. Using this result, we obtain that the minimum average sum AoI in a closed-form expression which can be interpreted as fundamental limit of sum AoI in a single relay network. Also, for error-prone and symmetric case, we have proved that greedy policy achieves the minimum average sum AoI at the destination nodes. For general case, we have proposed scheduling policy obtained via reinforcement learning.

cs.IT↗

arXiv2020

Wireless Distributed Edge Learning: How Many Edge Devices Do We Need?

We consider distributed machine learning at the wireless edge, where a parameter server builds a global model with the help of multiple wireless edge devices that perform computations on local dataset partitions. Edge devices transmit the result of their computations (updates of current global model) to the server using a fixed rate and orthogonal multiple access over an error prone wireless channel. In case of a transmission error, the undelivered packet is retransmitted until successfully decoded at the receiver. Leveraging on the fundamental tradeoff between computation and communication in distributed systems, our aim is to derive how many edge devices are needed to minimize the average completion time while guaranteeing convergence. We provide upper and lower bounds for the average completion and we find a necessary condition for adding edge devices in two asymptotic regimes, namely the large dataset and the high accuracy regime. Conducted experiments on real datasets and numerical results confirm our analysis and substantiate our claim that the number of edge devices should be carefully selected for timely distributed edge learning.

cs.DC↗

arXiv2020

Mobility-Aware Content Placement for Device-to-Device Caching Systems

User mobility has a large effect on optimal content placement in D2D caching networks. Since a typical user can communicate neighboring users who stay in the D2D communication area of the typical user, the optimal content placement should be changed according to the user mobility. Under consideration of randomness of incoming and outgoing users, we formulate an optimization problem to minimize the average data load of a BS. It is proved that minimization of the average data load of a BS can be transformed to maximization of a monotonic submodular function with a matroid constraint, for which a greedy algorithm can find near-optimal solutions. Moreover, when motions of neighboring users are rapid, the optimal content placement is derived in closed-form, aided by reasonable approximation and relaxation. In the high mobility regime, the optimal content placement is shown to cache partial amounts of the most popular contents.

cs.IT↗

arXiv2016

Optimal caching placement for wireless femto-caching network

This paper investigates optimal caching placement for wireless femto-caching network. The average bit error rate (BER) is formulated as a function of caching placement under wireless fading. To minimize the average BER, we propose a greedy algorithm finding optimal caching placement with low computational complexity. Exploiting the property of the optimal caching placement which we derive, the proposed algorithm can be performed over considerably reduced search space. Contrary to the optimal caching placement without consideration of wireless fading aspects, we reveal that optimal caching placement can be reached by balancing a tradeoff between two different gains: file diversity gain and channel diversity gain. Moreover, we also identify the conditions that the optimal placement can be found without running the proposed greedy algorithm and derive the corresponding optimal caching placement in closed form.

cs.IT↗