Abstract
Semantic Communication (SemCom) is a cornerstone of the 6G paradigm, shifting the focus from bit-level transmission to the delivery of underlying meaning. However, the performance of SemCom systems is often bottlenecked by the quality of the shared Knowledge Base (KB), which frequently contains "dirty" data such as mislabeled samples and class imbalances. This paper proposes the Task-Agnostic Learnable Weighted-Knowledge Base (TALSC) framework. By leveraging a meta-learning approach powered by Kolmogorov-Arnold Networks (KANs), TALSC dynamically evaluates the significance of KB samples. Our results demonstrate that this framework achieves over 12% higher accuracy than current state-of-the-art methods in noisy environments while maintaining task-agnostic flexibility.
1. Introduction
In the evolution toward 6G, Semantic Communication (SemCom) aims to optimize bandwidth by transmitting only the relevant features of data. This process relies on a Knowledge Base (KB) at both the transmitter and receiver to align semantic understanding. In real-world deployments, these KBs are rarely pristine; they suffer from Label Flipping Noise (incorrectly categorized data) and Class Imbalance (disproportionate representation of certain categories).
Existing systems often overfit to these biases, leading to significant performance degradation when faced with diverse intelligent services. This research introduces a robust, task-agnostic framework designed to handle imperfect data without requiring constant manual retraining or perfect KB synchronization.
2. The TALSC Framework
The proposed TALSC framework introduces a bi-level optimization strategy to manage KB quality. The system is divided into two primary components: the Learner (the semantic coding networks) and the Meta-Learner (the Sample Confidence Module, or SCM).
The SCM evaluates the "significance" of each sample in the KB by observing the feedback from the receiver. Instead of simply filtering out noisy data, the system learns a continuous weighting function. This allows the model to extract useful features even from imperfect samples while suppressing the negative influence of incorrect labels or over-represented classes.
3. Innovation: KANs and Significance Evaluation
A key innovation in TALSC is the use of Kolmogorov-Arnold Networks (KANs) for the SCM, replacing traditional Multi-Layer Perceptrons (MLPs). Unlike MLPs, which use fixed weights on nodes, KANs utilize learnable spline functions on the edges. This architectural choice allows the SCM to capture complex, non-linear significance functions more effectively.
As shown in the significance distributions, the SCM successfully distinguishes between reliable and unreliable data. For label flipping, it assigns low significance to high-loss (corrupted) samples. For class imbalance, it boosts the significance of minority class samples to ensure they are not ignored during the learning process. Furthermore, the authors introduced a Grid Extension (GE) method, allowing the KANs to increase their evaluation granularity (from coarse to fine) without retraining from scratch, significantly improving computational efficiency.
4. Experimental Results
The framework was rigorously tested using the CIFAR-10 dataset under various corruption scenarios:
- Label Noise: With a Flipping Noise Rate (FNR) of 0.4 (40% wrong labels), TALSC maintained high Semantic Recovery Accuracy (SRA), while the baseline TA-DeepJSCC saw a 65% drop.
- Class Imbalance: TALSC significantly improved the F1-score for minority classes, even with an imbalance factor as high as 800x.
- Efficiency: The KAN-based SCM achieved lower meta-loss than MLP-based alternatives. Using Grid Extension, the model could be fine-tuned in just 5 epochs, compared to the 80 epochs required for a full retraining cycle.
5. Conclusion and Future Work
TALSC provides a robust solution for semantic communications in the presence of imperfect knowledge bases. By dynamically weighting sample significance through a meta-learning KAN-based module, the system achieves superior accuracy and resilience. Future research will focus on reducing the dependency on small sets of "unbiased metadata" and testing the framework against more dynamic, non-stationary 6G channel models to ensure real-world viability.
Summary updated