DyCo-CL: Geometry-Aware Contrastive Learning Boosts Few-Shot Modulation Recognition
A novel self-supervised learning approach called Dynamic-Consistency Contrastive Learning (DyCo-CL) has demonstrated a 6.27% improvement in accuracy for 1-shot automatic modulation recognition (AMR) compared to earlier techniques. This method integrates Virtual Adversarial Augmentation (VAA) alongside a semantic consistency loss, serving as an implicit spectral regularizer to facilitate stable manifold exploration. Additionally, a Signal-Adaptive Swin Backbone with fixed-window attention limits the locality of attention, while a Hybrid Knowledge Fusion module reinforces representations using physical priors. Testing on RML benchmarks showcases the method's efficacy in few-shot scenarios.
Key facts
- DyCo-CL achieves 6.27% accuracy gain in 1-shot AMR
- Framework uses Virtual Adversarial Augmentation and semantic consistency loss
- Acts as implicit spectral regularizer for encoder
- Signal-Adaptive Swin Backbone with fixed-window attention improves structural stability
- Hybrid Knowledge Fusion module anchors representations with physical priors
- Experiments on RML benchmarks
- Addresses ineffective isotropic augmentations, spectral instability, and semantic drift
- Standard SSL for AMR struggles with these challenges
Entities
Institutions
- arXiv