Deep Learning for CSI Feedback

• Deep Learning for CSI Feedback is a method that uses neural network architectures to compress, quantize, and reconstruct wireless channel state information for efficient MIMO performance.
• It employs an encoder–decoder pipeline with trainable vector quantization and temporal difference strategies to minimize feedback overhead while preserving accuracy.
• Temporal refinement modules further enhance NMSE and throughput, delivering over 4x performance improvements compared to legacy feedback systems.

Deep learning for channel state information (CSI) feedback refers to the use of neural network methods to compress, quantize, report, and reconstruct channel state in wireless systems. In both next-generation Wi-Fi and 5G/6G cellular standards, efficient CSI feedback is crucial for enabling high spectral efficiency, optimal beamforming, and robust link adaptation. Deep learning-based architectures surpass classical codebook and compressive sensing approaches in both accuracy and bitrate efficiency and enable integration with temporal and spatial channel properties, nonlinearity, and system heterogeneity.

1. Neural Network Architectures for CSI Feedback

The canonical deep learning-based CSI feedback pipeline for Wi-Fi, as implemented in (Shin et al., 29 May 2025), deploys an encoder–decoder architecture to process angle-parameterized CSI, particularly the Givens decomposition parameters describing the left and right singular vector matrices. The encoder receives the angle parameter maps $\Phi\in\mathbb{R}^{2\times N_a\times N_c}$, processes them through a stack of 2D convolutional layers with Generalized Divisive Normalization (GDN) activations, followed by a fully connected layer to yield a compact latent code $\mathbf z\in\mathbb{R}^M$ (typically $M=256$).

To bridge the gap between continuous-valued representation and finite-rate feedback, a trainable vector quantization (VQ) module is introduced. Product VQ splits the latent code into $N$ subvectors, assigning each to a codeword in a shared codebook, yielding a quantized codebook index vector. The corresponding decoder mirrors the encoder’s structure in reverse, decompressing the quantized code and reconstructing the angle maps.

Unlike static traditional approaches, the encoder–decoder–VQ system is trained jointly end-to-end using loss functions adapted to the periodic nature of angle parameters. Reconstruction error is calculated as $$d(\Phi,\hat\Phi)=\|\exp(j\Phi)-\exp(j\hat\Phi)\|_F^2$$, ensuring invariance to angle wrap-around.

2. Leveraging Temporal Correlation for Feedback Efficiency

Modern Wi-Fi and wireless channels are temporally correlated, especially for pedestrian or indoor mobility. The framework in (Shin et al., 29 May 2025) exploits this structure through an "angle-difference" strategy. Rather than feeding back the entire $\Phi_t$ at every interval, the system computes a minimal, periodicity-aware difference

$\widetilde{\Delta}\Phi_t = h(\Phi_t - \Phi_{t-1}),$

where $h(x)$ outputs the smallest-magnitude rotation in $(-\pi,\pi)$. If the number of significant elements (above a threshold $\mu_{\rm th}$) in $\widetilde{\Delta}\Phi_t$ is below $N_{\rm th}$, only the difference is sent. Otherwise, the full angle map is retransmitted. One additional bit signals which mode ("diff" or "full") is used per interval.

To mitigate error propagation, two VQ structures are proposed:

• Parallel VQ: Two-stage quantizer encodes present information and the previous residual latent in separate stages, allowing refinement at the access point (AP).
• Unified VQ: A single quantizer operates on the sum of the current difference encoding and residuals, simplifying implementation.

Both variants support efficient delta-encoding with explicit error correction through regular full snapshots.

3. CSI Refinement via Temporal and Nonlinear Neural Networks

To further enhance reconstruction quality, a deep refinement module is deployed at the AP. This module, inspired by video prediction architectures (cf. SimVP, Fig. 9 in (Shin et al., 29 May 2025)), integrates several consecutive angle maps $\hat\Phi_{t-T+1},\dots,\hat\Phi_t$ and outputs a refined map $\check\Phi_t$. The architecture is a hybrid of 2D and 3D convolutional blocks, explicitly leveraging temporal redundancy.

Training proceeds in three phases:

1. Encoder–decoder–VQ is first trained on static feedback.
2. Angle-difference components and associated VQ are attached and jointly optimized.
3. The refinement network is added, with previous modules frozen, and trained via MSE loss in the angular domain.

Recursive training, in which the refinement net's own outputs are used as input context in further epochs, improves robustness and accuracy.

4. End-to-End Training Losses and Optimization

The system is trained end-to-end using a composite loss function: $$\mathcal L_{\text{VQ}} = d(\Phi,\hat\Phi) + \sum_i \|\text{sg}(\mathbf z_i) - \mathbf z_{q,i}\|_2^2 + \beta \sum_i \|\mathbf z_i - \text{sg}(\mathbf z_{q,i})\|_2^2,$$ where $\text{sg}(\cdot)$ applies stop-gradient, and $\beta \approx 0.25$ controls commitment. For the refinement module: $$\mathcal L_{\text{ref}} = \|\exp(j\Phi_t) - \exp(j\check\Phi_t)\|_F^2.$$ No extra dropout regularization is needed. Training uses Adam optimizer with standard learning rates and is conducted over thousands of epochs to convergence.

5. Quantitative Performance and Comparison to Legacy Systems

The framework is evaluated on DeepMIMO I3 Wi-Fi scenarios (2.4 GHz, 20 MHz BW, $8 \times 2$ MIMO, $N_c=64$). At a feedback overhead of 576 bits per report:

• Initial DL+VQ: NMSE $\approx -14.8$ dB, outperforming IEEE 802.11 standard ($-10$ dB). Net throughput is $51$ Mb/s vs $19$ Mb/s/$15$ Mb/s (legacy T0/T1).
• Angle-difference + unified VQ: Further NMSE improvements of $3$–$4$ dB (to $-18.7$ dB) and net throughput $58.9$ Mb/s.
• CSI refinement: Final NMSE sub $-23.8$ dB, net throughput $64.7$ Mb/s.

This demonstrates a greater than 4x throughput improvement over legacy feedback, with orders of magnitude reduced reporting overhead ($\sim$600 bits vs $10^4$ bits in standard IEEE 802.11 feedback).

6. Practical Considerations and Modular Extensions

The architectural complexity is moderate: encoder and decoder require $\sim$0.22M and $0.42$M parameters ($11$M and $28$M FLOPs per report, respectively). The refinement network incurs higher cost ($\sim$3M parameters, $200$M FLOPs for $T=3$) but resides only at the AP. All networks are compact enough ($\sim$MB) for embedded deployment.

The methodology is extensible to:

• Multi-user MIMO: through joint/groupwise angle-difference feedback.
• Frequency-selective OFDM: using sliding windows over subcarriers.
• Rate–distortion optimization: via entropy coding of VQ indices.
• Improved generalization: using meta-learning or explicit domain adaptation for deployment in untrained environments.

7. Significance in the Broader Context of CSI Feedback

The angle-parameter and temporal-difference approach in (Shin et al., 29 May 2025) represents a fully practical, standards-aware progression of DL-based CSI feedback, directly addressing the unique properties of Wi-Fi MIMO precoding matrices. The approach harmonizes high-fidelity feedback with tight overhead and computation budgets, setting the stage for rapid deployment in current and future Wi-Fi standards. The modular design—with encoder/decoder, quantization, temporal difference, and refinement—matches the architectural patterns found in the most performant contemporary DL-based feedback paradigms for both Wi-Fi and cellular systems, and establishes new benchmarks for normalized MSE and throughput in hardware-in-the-loop experiments (Shin et al., 29 May 2025, Qi et al., 8 Jul 2024).

---

Key References:

• Deep Learning-Based CSI Feedback for Wi-Fi Systems With Temporal Correlation (Shin et al., 29 May 2025)
• Deep Learning-based CSI Feedback in Wi-Fi Systems (Qi et al., 8 Jul 2024)