Choong Seon Hong’s research while affiliated with Kyung Hee University and other places

What is this page?


This page lists works of an author who doesn't have a ResearchGate profile or hasn't added the works to their profile yet. It is automatically generated from public (personal) data to further our legitimate goal of comprehensive and accurate scientific recordkeeping. If you are this author and want this page removed, please let us know.

Publications (41)


Computing Power in the Sky: Digital Twin-assisted Collaborative Computing with Multi-UAV Networks
  • Preprint

January 2025

·

2 Reads

Chao Wang

·

Yu Han

·

Long Zhang

·

[...]

·

Zhu Han

Resource-Efficient Federated Multimodal Learning via Layer-wise and Progressive Training

July 2024

·

8 Reads

Combining different data modalities enables deep neural networks to tackle complex tasks more effectively, making multimodal learning increasingly popular. To harness multimodal data closer to end users, it is essential to integrate multimodal learning with privacy-preserving training approaches such as federated learning (FL). However, compared to conventional unimodal learning, multimodal setting requires dedicated encoders for each modality, resulting in larger and more complex models that demand significant resources. This presents a substantial challenge for FL clients operating with limited computational resources and communication bandwidth. To address these challenges, we introduce LW-FedMML, a layer-wise federated multimodal learning approach, which decomposes the training process into multiple steps. Each step focuses on training only a portion of the model, thereby significantly reducing the memory and computational requirements. Moreover, FL clients only need to exchange the trained model portion with the central server, lowering the resulting communication cost. We conduct extensive experiments across various FL scenarios and multimodal learning setups to validate the effectiveness of our proposed method. The results demonstrate that LW-FedMML can compete with conventional end-to-end federated multimodal learning (FedMML) while significantly reducing the resource burden on FL clients. Specifically, LW-FedMML reduces memory usage by up to 2.7×2.7\times, computational operations (FLOPs) by 2.4×2.4\times, and total communication cost by 2.3×2.3\times. We also introduce a progressive training approach called Prog-FedMML. While it offers lesser resource efficiency than LW-FedMML, Prog-FedMML has the potential to surpass the performance of end-to-end FedMML, making it a viable option for scenarios with fewer resource constraints.


Fig. 2: Proposed AI framework for ENRA system.
Fig. 3: (a) Train and Validation performance. (b) Ground truth and predicted energy need data.
Fig. 4: Average scores for different number of EV allocation.
Demand Oriented Charging Resource Allocation for Electric Vehicles using Recurrent Neural Network
  • Conference Paper
  • Full-text available

April 2024

·

64 Reads

The rapid growth of electric vehicle (EV) demand in recent years requires a large number of charging stations and a huge amount of energy supply to facilitate the charging needs of EVs. Besides, EVs can recharge their batteries wirelessly by means of dynamic wireless charging while moving or stopped by a traffic signal. However, in a traffic situation, all the vehicles are not in need of energy, and the need for energy is not same for all the EVs. Therefore, it is necessary to analyze the energy needs of EVs in the present situation and allocate the charging resources to the needy EVs for dynamic wireless charging scenarios. In this work, we propose an energy-need-based resource allocation approach for dynamic charging of EVs using inductive power transfer methodology. An optimization problem is formulated to maximize the satisfaction scores of the needy EVs, which improves the charging efficiency of the system. Consequently, we propose an artificial intelligence framework to solve the formulated problem. We utilize a gated recurrent unit-based neural network model to predict battery energy needs and sorted the EVs in descending order for resource allocation according to these needs. The numerical results indicate that our proposed approach outperforms the baseline methods, which assures the effectiveness of our proposed system.

Download

Fig. 1. System model for an RIS-aided indoor VLC system.
When Visible Light Communication Meets RIS: A Soft Actor-Critic Approach

January 2024

·

30 Reads

This letter considers a reconfigurable intelligent surface (RIS)-aided indoor visible light communication system, where a mirror array-based RIS is deployed to assist the communication from a light-emitting diode (LED) to multiple user terminals (UTs). We aim to maximize the sum-rate in an entire serving period by jointly optimizing the orientation of the RIS reflecting unit, the time fraction for the UT, and the transmit power at the LED, subject to the communication and illumination intensity requirements. To solve this high-dimensional non-convex problem, we transform it as a constrained Markov decision process. Then, a soft actor-critic (SAC)-based deep reinforcement learning algorithm is proposed with the goal of maximizing both the average reward and the expected policy entropy. Simulation results prove the effectiveness of the proposed SAC-based joint optimization design in improving the sum-rate and long-term average reward.



Fig. 2. The top-1 average test accuracy (%) of FedLCA on the MNIST dataset in different communication rounds with the degree of skewness α = 0.5 and λ = 1.0. The legend in this figure states that "layer1" represents the output of layer1 and the input of layer2; "layer2" represents the output of layer2 and the input of layer3; "layer3" represents the output of layer3 and the input of layer4 (i.e. the classification layer).
Fig. 3. The top-1 average test accuracy (%) with our strategy on the MNIST dataset in different layers' output with the degree of skewness α = 0.5 and λ = 1.0. Our strategy, when combined with CKA measurement, is denoted as 'CKA'. Similarly, our strategy combined with MSE measurement is denoted as 'MSE', and combined with COS measurement is denoted as 'COS'.
Knowledge Distillation in Federated Learning: Where and How to Distill?

October 2023

·

213 Reads

·

5 Citations

Federated learning (FL) advances the field of distributed machine learning for facilitating the privacy-preserving management of edge devices and central servers. However, the majority of data among edge devices is non-IID (not Independent and Identically Distributed), making it challenging to achieve edge intelligence. In this work, we attempt to mitigate the above issue through knowledge distillation (KD) by sharing knowledge between the central server and edge devices. Specifically, we first investigate where (i.e. which feature layer) to conduct the distillation for knowledge sharing. We find that setting the feature layer to that before the classification head yields superior performance. Moreover, we investigate how to conduct the KD in terms of loss choices. We test various types of losses for enhancing the knowledge sharing and find that Center Kernel Alignment (CKA) achieves the best performance among the investigated loss metrics. Overall, this work sheds new light on where and how to perform KD in FL. Experimental results on MNIST and Fashion-MNIST demonstrate that our finding yields a performance gain of at least 4%.


A Complete Survey on Generative AI (AIGC): Is ChatGPT from GPT-4 to GPT-5 All You Need?

March 2023

·

5,539 Reads

·

2 Citations

As ChatGPT goes viral, generative AI (AIGC, a.k.a AI-generated content) has made headlines everywhere because of its ability to analyze and create text, images, and beyond. With such overwhelming media coverage, it is almost impossible for us to miss the opportunity to glimpse AIGC from a certain angle. In the era of AI transitioning from pure analysis to creation, it is worth noting that ChatGPT, with its most recent language model GPT-4, is just a tool out of numerous AIGC tasks. Impressed by the capability of the ChatGPT, many people are wondering about its limits: can GPT-5 (or other future GPT variants) help ChatGPT unify all AIGC tasks for 1 2 Zhang et al. diversified content creation? Toward answering this question, a comprehensive review of existing AIGC tasks is needed. As such, our work comes to fill this gap promptly by offering a first look at AIGC, ranging from its techniques to applications. Modern generative AI relies on various technical foundations, ranging from model architecture and self-supervised pretraining to generative modeling methods (like GAN and diffusion models). After introducing the fundamental techniques, this work focuses on the technological development of various AIGC tasks based on their output type, including text, images, videos, 3D content, etc., which depicts the full potential of ChatGPT's future. Moreover, we summarize their significant applications in some mainstream industries, such as education and creativity content. Finally, we discuss the challenges currently faced and present an outlook on how generative AI might evolve in the near future.



Fig. 1. System Model for Communication Cost Reduction for Semantic Communication.
Fig. 2. Encoder-Decoder LSTM-RNN Model for Text Summarization.
Fig. 3. Training loss vs. validation loss.
RNN-based Text Summarization for Communication Cost Reduction: Toward a Semantic Communication

January 2023

·

304 Reads

·

10 Citations

Text summarization has become a subject of great importance because of the continuous global growth of internet technology. The frequent occurrence of the same textual information creates an overhead in the communication network. Extracting the summary from a large document is quite challenging for any human being. Text summarization can play a vital role in this regard. To be precise, text summarization is the process of automatically producing and condensing the content of a given document into a more manageable form with a coherent message. Even though it only contains a few sentences, this concise explanation would nonetheless effectively convey the key idea. This illustrates how text summarization retains the essential information while substantially lowering the quantity of information that must be communicated. In this research, we propose a system model architecture where a central base station sends the summarized text to the users on their edge devices. We also demonstrate how the communication cost is reduced with each transmission of the condensed text. To ensure the best possible summary, we choose the long short-term memory recurrent neural network (LSTM-RNN) since RNN-based deep learning models have achieved great success in text analysis over the past few years. Additionally, the experimental results also show that our proposed model, in combination with LSTM-RNN reduces communication costs by an average of 85%.


A Goal-Oriented Semantic Communication Framework for Connected and Autonomous Vehicular Network: A Deep Auto-Encoder Approach

December 2022

·

349 Reads

·

2 Citations

Semantic communication will significantly increase data transmission effectiveness by only transferring the semantic features.Unfortunately, most of the study in this area primarily concentrates on a single type of application, such as image, text, oraudio. However, some of the existing studies also focus on goal-oriented communication. In this study, a goal-oriented semanticcommunication framework for the vehicular ad-hoc network has been proposed. A deep autoencoder (DAE) has been used tocapture the semantic information from traffic signs. The encoder of the DAE extracts the semantic information from a trafficsign, and then this semantic information is transmitted to the CAVs. After receiving the semantic information, the CAVs use thedecoder of the DAE to reconstruct the traffic sign. Then the CAVs use a Deep Q-Network (DQN) to take the appropriate actionbased on the reconstructed traffic sign. The experimental result indicates that our proposed model can minimize up to 90.81% ofthe communication cost.


Citations (26)


... GAN-based [3] and SMOTE-based [24] FL frameworks like BalanceFL [17] and FEDGAN-IDS [18] allow synthetic data generation at the client level to better represent minority classes. Knowledge distillation (KD) [15] in FL facilitates the transfer of knowledge from a complex model (teacher) to a simpler model (student), enhancing performance and efficiency. Federated Distillation (FD) addresses heterogeneity by supporting diverse model architectures, adapting to varying system capabilities, and facilitating knowledge transfer to overcome data distribution discrepancies, thereby enhancing performance [9]. ...

Reference:

Framework for Co-distillation Driven Federated Learning to Address Class Imbalance in Healthcare
Knowledge Distillation in Federated Learning: Where and How to Distill?

... On the server side, regularization is applied during the aggregation of model updates to prevent overfitting and improve model generalization. Techniques such as L1 and L2 regularization introduce penalty terms in the loss function during training, which helps to control the complexity of the shared model and ensures robust pattern learning from local data [133,134]. On the client side, participants implement regularization methods like dropout, batch normalization, or weight decay during their training processes [135,136]. ...

Federated Learning with Intermediate Representation Regularization
  • Citing Conference Paper
  • February 2023

... In [67], the authors proposed a SC system that can support speech to text transcription task and speech recovery task simultaneously. The differences and advantages of existing DL based methods for extracting SI from speech data are summarized in Table 3. • SI extracted from textual data: DL based text SC systems can be used for tasks such as text recovery [68]- [70] and speech to text summarization [71]. Different from images and speeches, textual data cannot be directly represented as numerical vectors or matrices that the learning model can process. ...

RNN-based Text Summarization for Communication Cost Reduction: Toward a Semantic Communication

... Moreover, a semantic forward mode is introduced to enable the relay node to transmit semantic information directly. In the study described in [23], a goal-oriented SC framework is proposed for VANETs. This framework utilizes a DAE to capture semantic information from traffic signs, which is then transmitted to connected autonomous vehicles. ...

A Goal-Oriented Semantic Communication Framework for Connected and Autonomous Vehicular Network: A Deep Auto-Encoder Approach

... The forthcoming 6G wireless communication networks are expected to provide widespread mobile connectivity, faster data services with lower power consumption, and seamless integration to accommodate the exponential increase in users across various applications [1]- [3]. While massive MIMO systems enhance network coverage through spatial diversity, they are hindered by the limitations of phased arrays, which rely on high-resolution phase shifters [4], [5]. ...

Artificial Intelligence Framework for Intelligent Omni-Surface Assisted Holographic MIMO using Sequential Neural Network Model

... With the remarkable advancements in the performance of edge smart devices, the majority of computing tasks can be efficiently performed at the edge. This has given rise to the paradigm of mobile edge computing (MEC), which is envisioned as the next generation of computing networks [1], [2]. Nonetheless, the collection of data from distributed clients presents risks and challenges, predominantly due to the privacy nature of the vast amount of data involved. ...

An Optimized Computation Offloading and Resource Allocation Strategy in Mobile Edge Computing Using Deep Reinforcement Learning
  • Citing Conference Paper
  • June 2022

... FL is studied in time-critical industrial applications with a massive quantity of sensitive data. FL enables Industrial Internet of Things (IIoT) devices to train and develop an intelligent framework for task scheduling [127]. Since estimating the exact execution time is hard, especially in the Internet of Vehicles (IoV), providing an optimal task scheduling algorithm is a substantial challenge. ...

Federated Learning Over the Industrial Internet of Things: A Joint Optimization of Edge Association and Resource Allocation

... The bi-directional auction driven by RL facilitates dynamic decision-making in the multi-agent markets. Yousafzai et al. [105] proposed to encourage/incentivize resourceconstrained entity's participation in collaborative model development. Likewise, Song et al. [106] proposed a marketplace that gives room for testing model efficiency/performance on a given use case before purchase. ...

FedMarket: A Cryptocurrency Driven Marketplace for Mobile Federated Learning Services

IEEE Access

... Furthermore, as the task load increases, it brings adverse effects on both system performance and UAV's battery lifetime [23]. UAV-enabled MEC systems, with UAVs acting as relays, can be used to improve user experience (based on service latency [9]) in IoT networks by offloading the computational tasks to more resourceful ground ESs [21], [24]. Unlike fixed relays, UAVs can quickly move to optimal locations, providing improved coverage and line-of-sight communication, especially in areas where buildings or infrastructure obstruct signal transmission. ...

Energy-Efficient Resource Allocation in Multi-UAV-Assisted Two-Stage Edge Computing for Beyond 5G Networks
  • Citing Article
  • September 2022

IEEE Transactions on Intelligent Transportation Systems

... Furthermore, the crucial hardware implementation of the quantization process itself remains unexplored, preventing an end-to-end acceleration. 2) Existing accelerator architectures fall into two categories: streaming-like [11,16,18,24,34,36] and processor-like [10,14,20,23,25,27,33,39,41], but neither offers an ideal solution for edge deployment with binarized Transformers. Streaming-like Alg Tech. ...

A Deep Learning Accelerator Based on a Streaming Architecture for Binary Neural Networks

IEEE Access