Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LOREN: Low Rank-Based Code-Rate Adaptation in Neural Receivers

Bram Van Bolderik, Vlado Menkovski, Sonia Heemstra de Groot, Manil Dev Gomony

TL;DR

LOREN addresses the memory and power burden of neural receivers that must support multiple code rates by freezing a shared base network and learning small, per-code-rate low-rank adapters. These adapters, applied via lightweight 1×1 updates, enable dynamic code-rate adaptation with negligible parameter overhead while achieving performance on par with or better than full retraining. End-to-end training over realistic 3GPP CDL channels demonstrates robust BLER performance across CRs, and hardware synthesis in 22nm FD-SOI shows significant area (over 65%) and power (up to 15%) reductions when supporting three CRs. The approach offers scalable, energy-efficient neural reception suitable for multi-CR wireless systems and paves the way for extending such adapters to broader modulation schemes and hyperparameter regimes.

Abstract

Neural network based receivers have recently demonstrated superior system-level performance compared to traditional receivers. However, their practicality is limited by high memory and power requirements, as separate weight sets must be stored for each code rate. To address this challenge, we propose LOREN, a Low Rank-Based Code-Rate Adaptation Neural Receiver that achieves adaptability with minimal overhead. LOREN integrates lightweight low rank adaptation adapters (LOREN adapters) into convolutional layers, freezing a shared base network while training only small adapters per code rate. An end-to-end training framework over 3GPP CDL channels ensures robustness across realistic wireless environments. LOREN achieves comparable or superior performance relative to fully retrained base neural receivers. The hardware implementation of LOREN in 22nm technology shows more than 65% savings in silicon area and up to 15% power reduction when supporting three code rates.

LOREN: Low Rank-Based Code-Rate Adaptation in Neural Receivers

TL;DR

LOREN addresses the memory and power burden of neural receivers that must support multiple code rates by freezing a shared base network and learning small, per-code-rate low-rank adapters. These adapters, applied via lightweight 1×1 updates, enable dynamic code-rate adaptation with negligible parameter overhead while achieving performance on par with or better than full retraining. End-to-end training over realistic 3GPP CDL channels demonstrates robust BLER performance across CRs, and hardware synthesis in 22nm FD-SOI shows significant area (over 65%) and power (up to 15%) reductions when supporting three CRs. The approach offers scalable, energy-efficient neural reception suitable for multi-CR wireless systems and paves the way for extending such adapters to broader modulation schemes and hyperparameter regimes.

Abstract

Neural network based receivers have recently demonstrated superior system-level performance compared to traditional receivers. However, their practicality is limited by high memory and power requirements, as separate weight sets must be stored for each code rate. To address this challenge, we propose LOREN, a Low Rank-Based Code-Rate Adaptation Neural Receiver that achieves adaptability with minimal overhead. LOREN integrates lightweight low rank adaptation adapters (LOREN adapters) into convolutional layers, freezing a shared base network while training only small adapters per code rate. An end-to-end training framework over 3GPP CDL channels ensures robustness across realistic wireless environments. LOREN achieves comparable or superior performance relative to fully retrained base neural receivers. The hardware implementation of LOREN in 22nm technology shows more than 65% savings in silicon area and up to 15% power reduction when supporting three code rates.
Paper Structure (13 sections, 2 equations, 8 figures, 1 table)

This paper contains 13 sections, 2 equations, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Background
  4. Proposed LOREN Architecture
  5. Base neural receiver architecture
  6. LOREN
  7. Training of the base network
  8. Training of the LOREN adapters
  9. Performance evaluation of LOREN
  10. Impact of LOREN rank and alpha
  11. Hardware implementation
  12. Power and area consumption of LOREN
  13. Conclusions and future work

Figures (8)

  • Figure 1: Overview of LOREN architecture that can dynamically adapt to varying CR with the help of layers that are augmented with LOREN adapters for each CR.
  • Figure 2: Overview of SIMO systems featuring a comparison between a classical receiver baseline, which employs a channel estimator with an option to use either exact CSI for an ideal theoretical performance or employing Least Squares (LS) estimation, and a neural receiverSionna, analyzed using a CDL_C channel model.
  • Figure 3: Overview of LOREN two-layer configuration that can dynamically adapt to varying Code rates(CR) with the help of layers that are augmented with per-CR LOREN adapters.
  • Figure 4: Loss plot for LOREN of QAM16, showing the changes in the loss value over the training iterations.
  • Figure 5: BLER plot depicting the performance of LOREN vs classical approaches and baselines. This shows that LOREN matches or outperforms classical approaches as well as least squares estimation.
  • ...and 3 more figures