Effect of batch normalization on quantization-aware training for deep neural network-based 5G NR channel estimation
Loading...
URL
Journal Title
Journal ISSN
Volume Title
School of Electrical Engineering |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
Department
Major/Subject
Mcode
Language
en
Pages
52
Series
Abstract
With the emergence of 5G New Radio (NR), wireless mobile communications have witnessed an increased rise in subscribers, enabling use cases that demand low latency, massive connectivity, and enhanced broadband. Channel estimation is an essential function that determines channel characteristics and ensures reliable signal detection.Data-driven methods such as deep learning have shown promising results for this task. However, the deployment of such methods on edge devices remains challenging due to resource and environmental constraints. Quantization is an approach that uses reduced precision to enable more efficient AI computation, which comes at the cost of accuracy degradation. In quantization, Quantisation-Aware Training (QAT) is a technique that trains neural networks with simulated quantization behaviour. Deep learning models often include batch normalization (BN) layers that stabilize and accelerate training, but managing these layers under QAT is non-trivial. The differences between training and inference behaviour of the layer necessitate proper layer folding, as incorrect handling of the layer can often degrade the performance. This thesis investigates the impact of batch normalization in the context of QAT for the channel estimation task. Based on the literature, different methods are selected and their effectiveness is evaluated in terms of NMSE under varying signal-to-noise ratio (SNR) conditions. The first part of results evaluates the feasibility and benefits of using batch normalization, where the floating-point model demonstrates significant improvement over both the least-squares baseline and the model without a BN layer. In the second part, three approaches are investigated and compared with a QAT baseline that performs static folding of BN parameters. The results show comparable performance at low SNRs, while improvements become evident at higher SNRs. The freeze-fold approach demonstrates the most promising results among the three methods, where BN parameters are frozen and folded after sufficient training.Description
Supervisor
Jäntti, RikuThesis advisor
Garifullin, AzatOlives, Jean-Luc