Abstract
In this work, we present the innovative Concrete Feedback Layers, designed to enable genuine bit-level, end-to-end Channel State Information (CSI) feedback using deep learning techniques. Overcoming the limitations of traditional discrete operations that impede gradient flow, these layers leverage the concrete distribution to facilitate efficient learning processes. Our extensive simulations reveal that these layers significantly enhance digital CSI feedback, achieving superior performance in terms of Normalized Mean Squared Error (NMSE) and cosine similarity compared to conventional feedback models. Furthermore, the integration of the Concrete Feedback Layers with the Feedback Bit Masking Unit (FBMU) allows for authentic bit-level variable-length CSI feedback, while maintaining a single adaptable model for various feedback lengths. This advancement marks a major leap forward in deep learning-based CSI feedback methods, potentially revolutionizing 6G communication systems with its flexibility and efficiency.
| Original language | English |
|---|---|
| Pages (from-to) | 15353-15366 |
| Number of pages | 14 |
| Journal | IEEE Transactions on Wireless Communications |
| Volume | 23 |
| Issue number | 10 |
| DOIs | |
| State | Published - 2024 |
Keywords
- 6G wireless systems
- Machine learning for communications
- channel feedback
- channel state information (CSI)
- deep learning
- end-to-end learning
- multiple-input multiple-output
- neural network architectures
- variable-length feedback