Autoencoder-Based CSI Compression

The encoder is a neural network residing on the user equipment (UE) that compresses the high-dimensional CSI matrix into a low-dimensional latent code vector. This process typically involves:

• Convolutional layers to exploit the spatial correlation of the antenna array.
• Fully connected layers to reduce dimensionality.
• Quantization of the latent code to a finite number of bits for digital feedback. The compression ratio (CR) can be extreme, often reducing a 2048-element vector to fewer than 64 real numbers.

The decoder is a mirror neural network at the base station (gNB) that reconstructs the original high-dimensional CSI matrix from the received low-dimensional latent code. Key design elements include:

• Transposed convolutions or upsampling layers to restore spatial resolution.
• Refinement blocks with residual connections to sharpen the reconstruction.
• Training to minimize the Normalized Mean Square Error (NMSE) between the original and reconstructed channel. The decoder effectively learns the manifold of realistic channel realizations.

The encoder and decoder are trained jointly as a single autoencoder using a dataset of channel realizations. This end-to-end optimization ensures the latent code captures the most salient features for reconstruction. The training process involves:

• Backpropagation through both the encoder and decoder simultaneously.
• A loss function combining reconstruction fidelity (e.g., MSE) and a sparsity or entropy constraint on the latent code.
• Simulating the quantization noise during training to ensure robustness to the limited feedback bit-budget.

To transmit the latent code over a finite-capacity feedback link, its continuous values must be quantized into discrete bits. Techniques include:

• Uniform scalar quantization with learnable step sizes.
• Vector quantization using a learned codebook, where the encoder output is mapped to the nearest codebook entry.
• Stochastic binarization during training, using a straight-through estimator to pass gradients through the non-differentiable quantization operation. The bit-width per latent dimension directly controls the trade-off between feedback overhead and reconstruction accuracy.

Modern autoencoder architectures employ multi-resolution or attention-based mechanisms to capture channel features at different scales:

• CsiNet+ uses a multi-rate convolutional structure to extract both fine-grained and coarse channel features.
• Attention mechanisms allow the network to dynamically focus on the most informative angular-delay domain paths.
• Transformer-based encoders treat the CSI matrix as a sequence of tokens, capturing long-range dependencies across the antenna array. This multi-scale processing is critical for handling diverse propagation environments.

A critical preprocessing step transforms the CSI from the spatial-frequency domain to the angular-delay domain using a 2D Discrete Fourier Transform (DFT). In this domain:

• The channel energy is concentrated in a small number of significant paths.
• The matrix becomes highly sparse, making it significantly easier for the autoencoder to compress.
• Truncation of the delay domain to the cyclic prefix length further reduces the input dimension before encoding. This transformation is the key enabler for achieving high compression ratios.

The complex-valued matrix capturing the amplitude and phase distortion a wireless signal experiences between every transmit-receive antenna pair. In massive MIMO systems with 64+ antennas, raw CSI feedback can consume 20-40% of uplink bandwidth. The CSI matrix is typically represented in the angular-delay domain where sparsity emerges—most energy concentrates in a few dominant paths—making it amenable to compression. Classical codebook-based approaches quantize this matrix, but autoencoders learn a non-linear manifold that preserves reconstruction fidelity at compression ratios exceeding 16:1.

The low-dimensional vector at the center of the autoencoder architecture where CSI is maximally compressed. The encoder network at the user equipment (UE) maps the high-dimensional CSI matrix into this compact representation, typically 8-32 floating-point values instead of thousands. The bottleneck forces the network to learn a disentangled manifold of channel features—separating path gains from angles of departure. Key design trade-offs include:

• Narrower bottlenecks achieve higher compression but risk losing multipath diversity
• Wider bottlenecks preserve reconstruction accuracy at the cost of feedback overhead
• Adding entropy coding after the latent code further reduces bitrate

The architectural split where the encoder runs on resource-constrained UE while the decoder executes at the computationally rich base station. This asymmetry is deliberate: the encoder must be lightweight enough for mobile handsets—often using depthwise separable convolutions or pruned architectures under 50K parameters—while the decoder can employ deeper residual networks with transposed convolutions for high-fidelity reconstruction. Practical considerations include:

• Encoder inference latency must stay under 1 ms on mobile DSPs
• Decoder can leverage batch normalization and wider layers without power constraints
• Quantization-aware training ensures the latent code survives 4-8 bit quantization for transmission

A paradigm where the autoencoder is trained jointly with the downstream task—beamforming vector computation or precoding matrix design—rather than minimizing raw reconstruction error. The loss function becomes the achievable sum-rate or signal-to-interference-plus-noise ratio (SINR) at the base station output. This task-aware training ensures the latent code preserves features critical for multi-user MIMO performance while discarding information irrelevant to beamforming. Techniques include:

• Differentiable beamforming layers appended after the decoder during training
• Multi-task learning balancing reconstruction fidelity with downstream throughput
• Adversarial training to ensure robustness against channel aging between feedback and transmission

Training strategies ensuring the continuous latent code survives discretization into a finite bitstream for actual uplink transmission. Without explicit quantization awareness, autoencoder reconstruction collapses when the latent vector passes through a uniform scalar quantizer with 4-8 bits per dimension. Mitigation approaches include:

• Straight-through estimator (STE) approximating the zero-gradient of quantization during backpropagation
• Soft-to-hard quantization annealing gradually sharpening the quantization grid during training
• Vector quantization (VQ) learning a discrete codebook of latent vectors, turning CSI feedback into a codebook index lookup
• Rate-distortion loss jointly optimizing reconstruction quality and entropy of the quantized latent code