Learnware for CSI Feedback: Scene-specific Small Models Can Do Big
| Source: arXiv AI
Tags: 6G, CSI feedback, learnware, model reuse, wireless communication, federated learning
A model repository framework for 6G base stations lets sites retrieve pre-trained CSI feedback models matched to their environment without transmitting raw channel data, achieving 18–58% accuracy gains over general models with up to 1,000x reduction in local fine-tuning.
Details
Channel State Information (CSI) feedback is the bottleneck for achieving 6G's spectral efficiency targets: large neural networks generalize across environments but are computationally expensive, while small scene-specific models need costly retraining for each new base station. This paper introduces a Learnware-based repository where a central AI data center stores scene-specific CSI models, each tagged with a specification combining network architecture parameters and a codebook-fingerprint embedding of its training distribution. A base station submits only its local fingerprint — not raw CSI — to retrieve the best matching pre-trained model, preserving data privacy. The retrieval strategy achieves over 90% selection accuracy. In simulation, the retrieved models improve on a general-purpose baseline by 18.8% in LOS scenarios and 57.7% in NLOS scenarios, while reducing local fine-tuning from thousands of samples to 1,000 samples and cutting training epochs by 100. The work is accepted by IEEE Transactions on Wireless Communications. This is a niche but practically relevant result for 6G infrastructure vendors. The privacy-preserving retrieval angle is commercially interesting since base stations won't share raw user data for model selection.