Implementation of Adapter-Based Transfer Learning for Channel Estimation Using a Pretrained Transformer in Multi-Channel Environments
- Authors
- Joo Won Lee, Kae Won Choi
- Venue
- Proceedings of Symposium of the Korean Institute of Communications and Information Sciences, 2025 KICS Summer Conference
- Published
- 20 June 2025
- Location
- Jeju, Korea
- Paper
- 32E-6
- Pages
- 981-982
- ISSN
- 2383-8302
- Official PDF
- KICS official PDF
- Link
- Google Scholar citation
- Keywords
- Channel estimation, adapter-based transfer learning, pretrained transformer, multi-channel environments
Summary
This implementation study adapts a Transformer channel estimator pretrained on InF-LoS, InF-NLoS, and RMa-LoS environments. It freezes the four Transformer encoder layers and trains inserted adapter modules separately for InF and RMa-LoS target environments, using NYUSIM-generated channels at 28 GHz.
Key Results
The reported NMSE comparison (Fig. 4) shows improved target-environment estimation accuracy after adapter tuning relative to the pretrained model in the evaluated InF (50 m) and RMa (300 m) settings.
Evaluation Scope
This is a simulation-based implementation study of environment-specific adapter transfer. It does not compare Adapter with LoRA or evaluate SNR-specific adapter switching; those questions are addressed in the separate ICTC and OJCOMS papers.
When to cite this work
Cite this work for an implementation example of adapter-based transfer learning with a frozen pretrained Transformer for multi-environment wireless channel estimation, especially InF and RMa scenarios. It provides evidence for environment-specific adapter tuning, rather than a general ranking of PEFT methods.
BibTeX
@inproceedings{lee2025adapterTransferKICS,
title = {Implementation of Adapter-Based Transfer Learning for Channel Estimation Using a Pretrained Transformer in Multi-Channel Environments},
author = {Lee, Joo Won and Choi, Kae Won},
booktitle = {Proceedings of Symposium of the Korean Institute of Communications and Information Sciences},
pages = {981--982},
year = {2025},
month = jun,
publisher = {Korean Institute of Communications and Information Sciences},
address = {Jeju, Korea},
issn = {2383-8302},
note = {2025 KICS Summer Conference, paper 32E-6},
url = {https://d2j16w31g89z0j.cloudfront.net/site/2025s/abs/1032-QKDHC.pdf}
}