BioLAMR: A Biomimetically Inspired Large Language Model Adaptation Framework for Automatic Modulation Recognition

Area of Science:

• Artificial Intelligence
• Signal Processing
• Wireless Communications

Background:

• Automatic modulation recognition (AMR) is crucial for adaptive wireless reception in communication-sensing systems.
• Existing AMR methods struggle at low signal-to-noise ratios (SNRs) and large language models (LLMs) face challenges with continuous I/Q signals.
• A modality gap exists between LLMs and continuous radio frequency (RF) signals.

Purpose of the Study:

• To develop a robust AMR framework, BioLAMR, that overcomes limitations of existing methods, particularly at low SNRs.
• To adapt LLMs for processing continuous I/Q signals by bridging the modality gap.
• To leverage bio-inspired processing inspired by the auditory system for improved feature extraction.

Main Methods:

• Proposed BioLAMR, a GPT-2 adaptation framework for AMR.
• Utilized a lightweight dual-domain fusion (LDDF) module for time- and frequency-domain feature extraction and fusion.
• Employed a convolutional embedding module to convert continuous I/Q signals into LLM-compatible sequences.
• Implemented a hierarchical fine-tuning strategy updating only 8.9% of parameters for efficient adaptation.

Main Results:

• BioLAMR achieved overall accuracies of 64.99% and 67.43% on RadioML2016.10a and RadioML2016.10b benchmarks, outperforming competitors.
• Under low-SNR conditions, BioLAMR reached 36.78% and 38.14% accuracy, demonstrating superior performance in challenging environments.
• Ablation studies confirmed the effectiveness of each component within the BioLAMR framework.

Conclusions:

• Combining dual-domain signal modeling with parameter-efficient LLM adaptation offers an effective strategy for robust AMR.
• BioLAMR demonstrates significant improvements in AMR performance, especially in low SNR conditions.
• The proposed framework successfully bridges the modality gap for LLMs processing continuous I/Q signals.