Wireless Resource Efficient Distributed Learning with Deep Joint Source-Channel Coding

Distributed learning methods such as federated learning and decentralized federated learning are promising approaches to constructing high-accurate models through collaboration among client nodes. In common distributed learning, each node continuously executes model training and sharing processes to...

Full description

Saved in:
Bibliographic Details
Published in:IEEE Globecom Workshops pp. 1 - 6
Main Authors: Matsumura, Issa, Suto, Katsuya
Format: Conference Proceeding
Language:English
Published: IEEE 08.12.2024
Subjects:
Accuracy
Computational modeling
Computer aided instruction
Data models
deep joint source-channel coding
deep learning
Distance learning
Distributed learning
Encoding
Noise
Quantization (signal)
Symbols
Wireless communication
wireless resource efficiency
ISSN:2166-0077
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Distributed learning methods such as federated learning and decentralized federated learning are promising approaches to constructing high-accurate models through collaboration among client nodes. In common distributed learning, each node continuously executes model training and sharing processes to sophisticate a trained model; however, it consumes numerous wireless resources because the data size of deep learning models has been increasing in recent years. For the remedy, this paper introduces a novel concept of distributed learning architecture that employs deep joint source-channel coding (DJSCC) for the sharing process. DJSCC can provide wireless resource efficient communications due to the data-driven encoder/decoder optimization and pseudo-analog modulation. The most significant contribution of this paper is that, even with noise included through pseudo-analog modulation, distributed learning smoothly trains models and achieves the desired accuracy with fewer transmission symbols. Through the computer simulation, we show that the proposal reduces the required symbols by approximately 90%, compared to the conventional distributed learning model.
ISSN:2166-0077
DOI:10.1109/GCWkshp64532.2024.11100602