---
title: 'CSI-SMoE: Sparse Experts for MIMO Transmission'
url: https://www.emergentmind.com/topics/csi-sparse-mixture-of-experts-csi-smoe
type: topic
---

# CSI-SMoE: Sparse Experts for MIMO Transmission

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 [2604.02691].

## 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** $f_{\rm dec}(\cdot)$: Recovers the image $\hat{\mathbf{x}}$ from $\hat{\mathbf{y}}$ and optionally $\mathbf{H}$.

The architectural novelty is the integration of an MoE “bottleneck” into the encoder and decoder networks, leveraging both real-time CSI features, $\phi_{\rm CSI}(\mathbf{H})$, and semantic features, $\phi_{\rm sem}(\mathbf{z})$, 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 $\mathbf{z}_m\in\mathbb{R}^{H_m\times W_m\times C}$ using two sublayers:

- **Window-based Multi-Head Self-Attention (W-MSA/SW-MSA):**
  $$
  \mathbf{u} = \mathrm{MSA}(\mathrm{LayerNorm}(\mathbf{z}_m)) + \mathbf{z}_m
  $$

- **Adaptive-MoE MLP:** Instead of a monolithic MLP, the design employs $E^s$ always-active shared experts and $E^r$ sparsely routed experts:
  $$
  \mathbf{w} = \frac{1}{E^s}\sum_{j=1}^{E^s} \mathrm{MLP}^s_j(\mathbf{v}) + \sum_{i\in\mathcal{A}} p_i\, \mathrm{MLP}^r_i(\mathbf{v})
  $$
  where $\mathbf{v} = \mathrm{LayerNorm}(\mathbf{u})$, $\mathcal{A}$ denotes the adaptively chosen routed-expert subset, and $p_i$ are routing probabilities.

Spatial resolution downsamples via layer-norm and fully connected "downsamplers" between stages. Patch embedding at the encoder input employs a $5\times 5$ 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:
$$
\phi_{\rm joint} = [\, \phi_{\rm sem}(\mathbf{v}) \;\|\; \phi_{\rm CSI}(\mathbf{H}) \,] \in \mathbb{R}^{d_c + d_h}
$$
- $\phi_{\rm sem}(\mathbf{v})$: Global average pooling of $\mathbf{v}$ followed by a fully connected layer.
- $\phi_{\rm CSI}(\mathbf{H})$: Preprocessed by flattening or singular-value summary of $\mathbf{H}$, then two FC layers.

A lightweight gating network produces scored logits $\mathbf{s} \in \mathbb{R}^{E^r}$ for $E^r$ routed experts, converted to selection probabilities via softmax.

**Threshold-based Top-K Routing:** Rather than a fixed $K$, active experts $\mathcal{A}$ are chosen by ranking the scores, initializing with the top score and accumulating further experts whose score gap does not exceed a threshold $T$, up to a maximum of $k$ 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: $L_b = E^r \sum_{i=1}^{E^r} f_i\,p_i$.
- Entropy regularizer: $L_e = -\frac{1}{B} \sum_{b=1}^B \sum_{i=1}^{E^r} p_{b,i}\,\log(p_{b,i}+\epsilon)$.
- Variance regularizer: $L_v = \sum_{i=1}^{E^r} (f_i-\tfrac1{E^r}\sum_j f_j)^2$.
Here, $f_i$ 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: $[2,2,6,2]$
- Channel dimensions: $[128,192,256,320]$
- Experts: shared $E^s=1$, routed $E^r=5$, max $k=2$, threshold $T=0.1$

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

The main training loss is the mean squared error (MSE):
$$
L_{\rm MSE} = \mathbb{E}\|\mathbf{x}-\hat{\mathbf{x}}\|^2_2
$$
An optional perceptual/LPIPS loss $L_{\rm perc}$ can be included. The composite objective is:
$$
L = L_{\rm MSE} + \lambda L_b + \alpha L_e + \beta L_v + \gamma L_{\rm perc}
$$
with $\lambda$, $\alpha$, $\beta$, $\gamma$ tuned by cross-validation.

## 5. Training Protocol and Implementation

Training is performed end-to-end with the Adam optimizer (initial learning rate $10^{-4}$) using 900 images from DIV2K for training and 24 $256\times256$ crops from Kodak for evaluation. Fading channel realizations are sampled for each mini-batch. CSI preprocessing entails splitting real/imaginary parts, flattening to a $2N_rN_t$ vector, and extracting features with two FC layers. The Swin window size is fixed at $8$.

## 6. Quantitative Evaluation and Ablation

Performance is benchmarked against DeepJSCC and SwinJSCC baselines for $2\times2$ and $8\times8$ MIMO. CSI-SMoE demonstrates PSNR gains of $0.8$–$1.5$ dB and LPIPS reductions of $5$–$15\%$ relative to SwinJSCC at equivalent bandwidth and SNR. At $R=0.0833$ (Kodak, $8\times8$ 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 $10$ dB ($R=0.0833$) 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 $1.68$ 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 [2604.02691].

Source: https://www.emergentmind.com/topics/csi-sparse-mixture-of-experts-csi-smoe