CSI-SMoE: Sparse Experts for MIMO Transmission

• The paper demonstrates that integrating a sparse MoE in both encoder and decoder backbones improves PSNR by up to 1.5 dB over traditional models.
• CSI-SMoE dynamically selects experts using real-time channel state and semantic features, providing tailored processing for MIMO fading channels.
• Empirical results indicate significant LPIPS reductions and robust performance under variable SNR and bandwidth constraints.

The CSI-Sparse Mixture-of-Experts (CSI-SMoE) framework is an adaptive semantic communication system designed for end-to-end wireless image transmission over multi-input multi-output (MIMO) fading channels. The primary innovation of CSI-SMoE is the integration of a sparse Mixture-of-Experts (MoE) architecture within both the semantic encoder and decoder backbones. Expert selection in this architecture is governed jointly by real-time channel state information (CSI) and semantic features of image patches, yielding a communication system that adapts its internal processing to both instantaneous channel and content variations. This dual-driven routing strategy addresses the rigidity and limited robustness inherent in conventional single-driven or fixed-model semantic communication methods (Wan et al., 3 Apr 2026).

1. End-to-End System Model and Notation

CSI-SMoE operates by transmitting an RGB image $\mathbf{x}\in\mathbb{R}^{H\times W\times 3}$ over an $N_t\times N_r$ MIMO fading channel with an average power constraint. The system consists of three key stages:

1. Semantic Encoder $f_{\rm enc}(\cdot)$: Maps $\mathbf{x}$ to a complex-valued matrix of transmit symbols $\mathbf{y}\in\mathbb{C}^{N_t\times K}$, with $K$ channel uses. The bandwidth ratio is defined as $R=\frac{K}{H\,W\,3}$.
2. MIMO Fading Channel: Modeled as $$\hat{\mathbf{y}} = \mathbf{H}\mathbf{y} + \mathbf{n}$$, where $\mathbf{H}\in\mathbb{C}^{N_r\times N_t}$ is the instantaneous CSI and $$\mathbf{n}\sim\mathcal{CN}(\mathbf{0},\sigma^2\mathbf{I})$$ is additive complex Gaussian noise.
3. Semantic Decoder $N_t\times N_r$0: Recovers the image $N_t\times N_r$1 from $N_t\times N_r$2 and optionally $N_t\times N_r$3.

The architectural novelty is the integration of an MoE “bottleneck” into the encoder and decoder networks, leveraging both real-time CSI features, $N_t\times N_r$4, and semantic features, $N_t\times N_r$5, for dynamic expert selection.

2. Adaptive MoE Swin Transformer Architecture

The core architectural unit is the Adaptive-MoE Swin-Transformer (AD–MoE ST) block. Each block processes patch-level features $N_t\times N_r$6 using two sublayers:

• Window-based Multi-Head Self-Attention (W-MSA/SW-MSA):

$N_t\times N_r$7

• Adaptive-MoE MLP: Instead of a monolithic MLP, the design employs $N_t\times N_r$8 always-active shared experts and $N_t\times N_r$9 sparsely routed experts:

$f_{\rm enc}(\cdot)$0

where $f_{\rm enc}(\cdot)$1, $f_{\rm enc}(\cdot)$2 denotes the adaptively chosen routed-expert subset, and $f_{\rm enc}(\cdot)$3 are routing probabilities.

Spatial resolution downsamples via layer-norm and fully connected "downsamplers" between stages. Patch embedding at the encoder input employs a $f_{\rm enc}(\cdot)$4 convolution of stride 2.

3. Dynamic Expert Gating and Routing Mechanism

Expert routing within each AD-MoE MLP is determined by a dynamic gating mechanism that synthesizes semantic and CSI feature vectors:

$f_{\rm enc}(\cdot)$5

• $f_{\rm enc}(\cdot)$6: Global average pooling of $f_{\rm enc}(\cdot)$7 followed by a fully connected layer.
• $f_{\rm enc}(\cdot)$8: Preprocessed by flattening or singular-value summary of $f_{\rm enc}(\cdot)$9, then two FC layers.

A lightweight gating network produces scored logits $\mathbf{x}$0 for $\mathbf{x}$1 routed experts, converted to selection probabilities via softmax.

Threshold-based Top-K Routing: Rather than a fixed $\mathbf{x}$2, active experts $\mathbf{x}$3 are chosen by ranking the scores, initializing with the top score and accumulating further experts whose score gap does not exceed a threshold $\mathbf{x}$4, up to a maximum of $\mathbf{x}$5 experts. This makes the number of active experts both input and CSI dependent.

To encourage load balancing and prevent expert collapse, three training-time MoE regularizers are incorporated:

• Load-balance loss: $\mathbf{x}$6.
• Entropy regularizer: $\mathbf{x}$7.
• Variance regularizer: $\mathbf{x}$8. Here, $\mathbf{x}$9 denotes empirical activation frequency.

4. Semantic Encoder/Decoder Design and Loss Function

The semantic encoder backbone comprises four stages: convolutional patch embedding, two Swin Transformer stages, two AD-MoE Swin-Transformer stages, and a final fully connected normalization layer. Hyperparameters follow:

• Stages: $\mathbf{y}\in\mathbb{C}^{N_t\times K}$0
• Channel dimensions: $\mathbf{y}\in\mathbb{C}^{N_t\times K}$1
• Experts: shared $\mathbf{y}\in\mathbb{C}^{N_t\times K}$2, routed $\mathbf{y}\in\mathbb{C}^{N_t\times K}$3, max $\mathbf{y}\in\mathbb{C}^{N_t\times K}$4, threshold $\mathbf{y}\in\mathbb{C}^{N_t\times K}$5

The decoder is architecturally symmetric and reuses weights for AD-MoE Swin Transformer blocks.

The main training loss is the mean squared error (MSE):

$\mathbf{y}\in\mathbb{C}^{N_t\times K}$6

An optional perceptual/LPIPS loss $\mathbf{y}\in\mathbb{C}^{N_t\times K}$7 can be included. The composite objective is:

$\mathbf{y}\in\mathbb{C}^{N_t\times K}$8

with $\mathbf{y}\in\mathbb{C}^{N_t\times K}$9, $K$0, $K$1, $K$2 tuned by cross-validation.

5. Training Protocol and Implementation

Training is performed end-to-end with the Adam optimizer (initial learning rate $K$3) using 900 images from DIV2K for training and 24 $K$4 crops from Kodak for evaluation. Fading channel realizations are sampled for each mini-batch. CSI preprocessing entails splitting real/imaginary parts, flattening to a $K$5 vector, and extracting features with two FC layers. The Swin window size is fixed at $K$6.

6. Quantitative Evaluation and Ablation

Performance is benchmarked against DeepJSCC and SwinJSCC baselines for $K$7 and $K$8 MIMO. CSI-SMoE demonstrates PSNR gains of $K$9–$R=\frac{K}{H\,W\,3}$0 dB and LPIPS reductions of $R=\frac{K}{H\,W\,3}$1–$R=\frac{K}{H\,W\,3}$2 relative to SwinJSCC at equivalent bandwidth and SNR. At $R=\frac{K}{H\,W\,3}$3 (Kodak, $R=\frac{K}{H\,W\,3}$4 MIMO), key PSNR results are:

SNR (dB)	DeepJSCC	SwinJSCC	CSI-SMoE
0	24.3	26.1	27.2
5	25.7	27.4	28.6
10	26.8	28.3	29.6
15	27.4	28.9	30.1

An ablation comparing routing signal sources at $R=\frac{K}{H\,W\,3}$5 dB ($R=\frac{K}{H\,W\,3}$6) demonstrates joint content-CSI gating achieves the best PSNR:

Routing Signal	PSNR (dB)
Content-Only	28.3
CSI-Only	28.9
Joint (Ours)	29.6

Analysis of expert activation frequencies shows on average $R=\frac{K}{H\,W\,3}$7 routed experts are active per patch with all five experts utilized, validating the balancing regularizers.

7. Impact and Significance

CSI-SMoE jointly leverages instantaneous CSI and semantic content for sparse expert routing in each Transformer block, dynamically adapting both the number and identity of active experts through a data- and channel-driven Top-K mechanism. This flexible architecture overcomes the rigid coupling and limited adaptation of prior MoE-based or single-driven systems. Empirical results establish substantial improvements in PSNR and perceptual metrics at constant bandwidth and compute budget, positioning CSI-SMoE as an advanced solution for adaptive, efficient, and robust wireless image semantic communication under time-varying channels (Wan et al., 3 Apr 2026).