MAP-Based Task-Oriented Precoding for Multiuser Communication

Technical Summary: MAP-Based Task-Oriented Precoding for Multiuser Communication

Problem Statement
The paper addresses the limitations of existing task-oriented and semantic communication frameworks in distributed wireless classification systems. While conventional systems focus on minimizing signal distortion or maximizing mutual information, these objectives are not explicitly aligned with end-task performance (e.g., classification accuracy). Existing learning-based approaches, such as Deep Joint Source–Channel Coding (JSCC) or methods utilizing the Maximal Coding Rate Reduction (MCR²) criterion, often rely on surrogate objectives that require computationally expensive operations, including repeated covariance matrix inversions and eigen-decompositions. Furthermore, these methods may exhibit sensitivity to hyperparameters and fail to directly optimize for the Maximum A Posteriori (MAP) decision rule under channel impairments.

Methodology
The authors propose a unified framework for distributed multiuser communication that jointly optimizes feature extraction and linear precoding based on an approximate MAP error formulation. The system consists of $K$ edge devices transmitting features to a central fusion center over a wireless channel.

1. System Model: The system models the transmission of feature vectors $x_k$ extracted from high-dimensional observations $s_k$ via a neural network $f_\theta(\cdot)$. These features are precoded by matrices $V_k$ and transmitted over a Rayleigh fading channel to a server with $M$ receive antennas.
2. MAP-Based Design:
   • Receiver Approximation: Assuming class-conditional Gaussian feature distributions, the authors derive a tractable MAP detector. To avoid the computational burden of inverting the full covariance matrix $K_j$, they employ an approximation where the covariance is treated as a scaled identity matrix ($K_j \approx \delta_H I_{MT}$). This simplifies the detector to a distance-based metric dependent on the class means.
   • Error Bound: The classification error probability is upper-bounded using a pairwise union bound involving the $Q$-function of the distance between class means in the received signal space.
3. Transmitter Optimization:
   • Feature Extraction: A neural network is trained to minimize a loss function derived from the pairwise error bound under a nominal noiseless channel assumption. This loss promotes the separation of class-dependent feature means ($\mu_j$) without requiring covariance estimation.
   • Precoder Design: The precoding matrix $V$ is optimized to minimize the pairwise MAP error approximation under a transmit power constraint. The non-convex $Q$-function is approximated by an exponential surrogate, allowing the use of a projected gradient descent scheme to update the precoder.
4. Complexity Reduction: The proposed approach explicitly avoids the $O(BD^2 + CD^3)$ complexity associated with MCR²-based methods (which require covariance estimation and log-determinant calculations) and the $O(D^3 + CM^3T^3)$ complexity of block coordinate ascent methods used in other precoders. Instead, the proposed method relies on pairwise mean separations, resulting in a complexity of $O(DC^2)$ for feature learning and $O(C^2 N_k K T (D_k + M))$ for precoding.

Key Contributions

• Direct Task Alignment: The framework directly targets post-channel MAP decision performance rather than relying on indirect reconstruction or mutual information objectives.
• Tractable Optimization: By approximating the channel-affected covariance matrix, the authors derive a design that eliminates the need for repeated covariance inversions and eigen-decompositions.
• Joint Optimization: The paper presents a coherent design for both the learning-based feature extractor and the linear precoder, optimized under a unified MAP-driven objective.
• Computational Efficiency: The proposed method significantly reduces computational complexity compared to state-of-the-art task-oriented schemes like MCR² and LMMSE-based precoders.

Numerical Results
Simulations were conducted using the CIFAR-10 dataset with a distributed setup involving two views per sample and ResNet-18 backbones.

• Feature Extraction: The proposed feature extractor achieved test accuracy and convergence behavior comparable to the MCR² method [8] across training epochs. However, unlike MCR², which showed strong sensitivity to the hyperparameter $\epsilon$ (with optimal performance at $\epsilon=5$ and degradation otherwise), the proposed method demonstrated robustness without sensitive tuning parameters.
• Precoding Performance: The proposed precoder achieved higher classification accuracy than both the MCR² precoder and the LMMSE baseline [5] across varying numbers of receive antennas.
• Complexity: The results confirmed that the proposed method achieves superior performance while simultaneously reducing computational complexity by avoiding expensive matrix operations.

Significance and Claims
The paper claims that its primary significance lies in bridging the gap between learning-based feature representation and channel-aware signal processing through a unified, low-complexity framework. By deriving a tractable class-mean separation objective, the authors demonstrate that it is possible to improve class separability after channel distortion without the computational overhead of covariance-based methods. The work asserts that this approach offers a more stable and efficient alternative for next-generation task-oriented wireless systems, achieving higher classification accuracy with reduced computational cost compared to existing schemes.