Task-Agnostic Weighted-KB Semantic Communication

• TALSC is a paradigm that integrates meta-learning with adaptive weighted sample evaluation for robust, task-agnostic semantic communications under noisy and imbalanced conditions.
• It combines semantic coding networks with a learnable weighted-knowledge base and a sample confidence module that dynamically reweights samples based on loss evaluations.
• The system achieves over 12% improvements in semantic recovery accuracy and image quality metrics, demonstrating effectiveness across diverse channels and real-world biases.

Task-Agnostic Learnable Weighted-Knowledge Base Semantic Communication (TALSC) is a paradigm within semantic communications specifically designed to enhance robustness, generalization, and sample efficiency in the face of data heterogeneity and real-world biases—such as label flipping noise and class imbalance—by integrating learnable, meta-optimizable knowledge bases with adaptive sample weighting strategies. TALSC achieves these goals by coupling traditional semantic coding networks with a learnable weighted-knowledge base (LW-KB) and a meta-learning-derived sample confidence module (SCM), resulting in a communication infrastructure that is task-agnostic, robust to unknown distribution shifts, and applicable across diverse scenarios and channel conditions (Jiang et al., 15 Sep 2025).

1. Architectural Foundations and Key Components

TALSC is architected around three interdependent modules:

• Semantic Coding Networks (“Learners”): Deep networks (e.g., CNNs), responsible for encoding raw data $S$ into channel inputs $x$ at the transmitter and reconstructing $S$ from received channel outputs at the receiver, optimized for tasks such as classification or image recovery.
• Learnable Weighted-Knowledge Base (LW-KB): A memory structure storing empirical samples $K = \{k_i\}$, each annotated with a dynamically learned significance weight $v_i$. This knowledge base is iteratively updated to reflect the empirical distribution and contribution of both clean and biased samples during training.
• Sample Confidence Module (SCM, Meta-Learner): An extra network that maps task loss feedback $L_i^{(\phi)}$—computed for each sample after pragmatic evaluation at the receiver—into a continuous significance value $v_i \in [0, 1]$. The SCM is responsible for adaptively downweighting misleading samples (e.g., label-flipped or minority-class instances) and upweighting reliable ones, using a significance evaluation function (SEF) implemented either as a multilayer perceptron or a Kolmogorov-Arnold network (KAN) for advanced nonlinear mapping and higher resolution scaling.

The training process operates in a bi-level structure: inner optimization updates the semantic coding networks using a weighted task loss, and outer meta-optimization refines the SCM parameters based on performance on unbiased meta-data, enabling sample reweighting to align with task objectives even as distributional properties evolve.

2. Meta-Learning Driven Adaptation and Bi-Level Optimization

The meta-learning aspect of TALSC is realized via the SCM’s bi-level optimization. At each training step, the learner weights $\omega$ are updated on the full (possibly noisy or imbalanced) training set using loss gradients weighted by the SCM output:

$$\omega^* = \omega - \alpha \frac{1}{n} \sum_{i=1}^n \mathcal{V}_i(\Theta) \cdot \nabla_{\omega} L^{(\phi)}_i(\omega)$$

where $\mathcal{V}_i(\Theta)$ is the output of the SCM ($\Theta$ denotes its parameters), and $L^{(\phi)}_i$ is the task loss for sample $i$. The SCM parameters are in turn meta-optimized to minimize the meta-loss over an unbiased batch $B$:

$$\Theta^{(t+1)} = \Theta^{(t)} - \beta \frac{1}{m} \sum_{j=1}^m \nabla_{\Theta} L^{(\phi)}_j(\omega^*(\Theta^{(t)}))$$

This closed-form expression (see Eq. (update-eq) in (Jiang et al., 15 Sep 2025)) ensures that loss-gradient alignment between base and meta-data accelerates the adaptation of the weighting function, steering learning away from bias or corrupted regions.

The SCM can use a standard MLP as SEF or, for improved flexibility, a KAN-based architecture. The KAN uses spline-based activation for higher-order nonlinear approximations of the mapping from loss to sample weight and supports grid extension (SCM-GE) for fine-grained post-training refinement, leveraging an analytical basis transformation (see Eq. (16)-(18) in (Jiang et al., 15 Sep 2025)) for scalable, resolution-adaptable significance evaluation without retraining.

3. Robustness to KB Bias: Label Flipping and Class Imbalance

A central design criterion is robustness under real-world KB bias. The TALSC framework achieves this via adaptive sample weighting and selective attention to empirically valuable data. In the presence of label flipping noise, as parameterized by a flipping noise rate (FNR), TALSC suppresses the influence of high-loss (likely corrupted) samples by assigning them low significance values. During class imbalance, the SCM adaptively increases the weights of underrepresented samples, as evidenced by improved F1-scores for minority classes even when majority/minority splits reach ratios as extreme as 9:1.

Confusion matrices confirm that the learner, guided by the SCM’s reweighting, maintains high discriminability for all classes despite severe class imbalance or noise, whereas fixed-weighted or non-adaptive baselines rapidly degrade.

4. Mathematical Foundations and Performance Metrics

TALSC introduces several formal tools and metrics:

• Weighted Task Loss: For sample $i$,

$$L^{(\phi)}(s_i, z_i; \omega) = (1-\lambda) \ell_{CE}(\phi(\hat{s}_i), z_i) + \gamma \lambda \ell_{MSE}(\hat{s}_i, s_i)$$

where $\lambda$ reflects compression rate and $\gamma$ is an MSE scaling factor.

• Weighted Learner Update:

$$\omega^* = \underset{\omega}{\arg\min}\, \frac{1}{N} \sum_{i=1}^N v_i \cdot L^{(\phi)}_i(\omega)$$

• KAN Approximation Bound: For KAN-based SEF, the approximation error between the true and estimated significance mapping $\mathcal{V}(\cdot)$ satisfies

$$\|\mathcal{V}(x) - \mathcal{V}(x; \Theta)\|_{C^m} \leq C G^{-k-1+m}$$

where $G$ is the grid size and $k$ the spline order, providing theoretical convergence guarantees during SCM-GE.

• Semantic Recovery Accuracy (SRA): Fraction of correct semantic recoveries (e.g., classification accuracy) at the receiver.
• Multi-Scale Structural Similarity (MS-SSIM): Image-based metric quantifying the preservation of structural information in reconstructed images.

Experimental results report SRA and MS-SSIM improvements of at least 12% over state-of-the-art semantic communication frameworks under both AWGN and fading channel conditions.

5. SCM-Grid Extension via Kolmogorov-Arnold Networks

The SCM-Grid Extension (SCM-GE) mechanism enables dynamic refinement of the significance evaluation granularity. After initial KAN training on a coarse grid, spline coefficients can be optimally transformed to a finer grid via a pseudoinverse-derived transformation matrix, as detailed in Eq. (19) of (Jiang et al., 15 Sep 2025). This technique yields an SEF with higher approximation accuracy and more detailed responsiveness to observed task loss distributions, all without requiring full retraining—critical for scalable deployment in heterogeneous 6G environments.

This approach provides controllable tradeoffs between computational complexity and evaluation precision, supporting robust operation even as KB content or task definitions evolve.

6. System Operation and Empirical Results

The operational flow of TALSC comprises data encoding, significance estimation, KB weighting, and semantic recovery. During system use:

1. The transmitter encodes data $S$ via $f(\cdot)$; the receiver applies $g(\cdot)$ and $\phi(\cdot)$ to obtain outputs.
2. The receiver computes the loss for each sample, passes these losses through the SCM’s SEF, and updates the importance weights in the LW-KB.
3. These weights are used to rescale the sample losses in the next round of learner updates, via the weighted loss minimization formula.
4. The SCM itself is periodically updated by meta-optimizing against a held-out meta-data set to maximize robustness and adaptability.

Simulation studies show that, even under 40% label flipping noise or severe class imbalance, the SRA and perceptual quality (MS-SSIM) of TALSC remain superior to competing methods (e.g., TA-DeepJSCC). This is attributed to the system’s dynamic sample evaluation and weighting strategy, as well as the meta-learning-optimized interaction between SCM and learner.

7. Significance, Applications, and Future Directions

The TALSC framework represents a methodologically rigorous, empirically validated strategy for robust semantic communication in open and highly variable environments. Its architecture is suitable for 6G scenarios with heterogeneous KBs, distributed IoT deployments, and scenarios requiring reliable machine intelligence at the network edge. Key contributions include:

• End-to-end learnability and adaptation to unknown task requirements (“task-agnosticity”).
• Robustness to label noise and imbalanced data through meta-learned sample weighting.
• Scalability and granularity control in its SCM via the grid extension and KAN approach.
• Empirically validated improvements—≥12% SRA and MS-SSIM gains—across a variety of channel models and task regimes.

These design elements position TALSC as a foundational solution for next-generation semantic communication systems confronting real-world biases and dynamic information landscapes (Jiang et al., 15 Sep 2025).