Zero-Shot Knowledge Base Resizing for Rate-Adaptive Digital Semantic Communication

Digital semantic communication systems, which often leverage the Vector Quantized Variational Autoencoder (VQ-VAE) framework, are pivotal for future wireless networks. In a VQ-VAE-based semantic communication system, the transmission rate is directly governed by the size of a discrete codebook known as knowledge base (KB). However, the KB size is a fixed hyperparameter, meaning that adapting the rate requires training and storing a separate model for each desired size – a practice that is too computationally and storage-prohibitive to achieve truly granular rate control. To address this, we introduce a principled, zero-shot KB resizing method that enables on-the-fly rate adaptation without any retraining. Our approach establishes a global importance ranking for all vectors within a single, large parent KB by uncovering its inherent semantic hierarchy. This is achieved via a three-step framework: 1) embedding KB vectors into hyperbolic space to reveal their hierarchical relationships; 2) constructing a master semantic tree using a minimum spanning tree algorithm; 3) enabling instant resizing by iteratively pruning the least important leaf nodes. Extensive simulations demonstrate that our method achieves reconstruction quality nearly identical to that of dedicated KBs trained from scratch, while demanding only a fraction of the computational budget. Moreover, our approach exhibits superior robustness at very low rates, where conventional KBs suffer from catastrophic failure. Our work resolves a fundamental limitation of VQ-VAE-based semantic communication systems, offering a practical and efficient path toward flexible and rate-adaptive semantic communication.

💡 Research Summary

The paper tackles a fundamental limitation of digital semantic communication systems built on vector‑quantized variational autoencoders (VQ‑VAEs). In such systems a shared discrete codebook, called the knowledge base (KB), determines both the transmission rate (through the number of bits needed to send an index) and the reconstruction quality (through quantization error). Traditionally the KB size K is a fixed hyper‑parameter set during training; adapting the rate therefore requires training and storing a separate model for each desired K, which is computationally and storage‑prohibitive.

To overcome this, the authors propose a zero‑shot KB resizing framework that can generate a smaller, rate‑appropriate KB from a single, large, pre‑trained KB without any retraining. The method consists of three main steps:

1. Hyperbolic Embedding – The vectors of the large KB are mapped from Euclidean space into the Poincaré ball using the exponential map. In hyperbolic space, distance from the origin naturally reflects semantic granularity: vectors near the origin correspond to coarse, generic concepts, while those near the boundary represent fine‑grained concepts.

2. Master Semantic Tree Construction – Using hyperbolic distances as edge weights, a minimum spanning tree (MST) is built with Prim’s algorithm. The node closest to the origin becomes the root, representing the most general concept. The MST reveals an implicit hierarchical structure among the codebook vectors.

3. Hierarchical Pruning – To obtain a KB of any target size K < Ķ, leaf nodes (the most fine‑grained concepts) are iteratively removed from the MST until exactly K nodes remain. This deterministic pruning preserves the foundational skeleton of the codebook while discarding the least essential vectors.

Finally, the retained hyperbolic vectors are projected back to Euclidean space via the logarithmic map, yielding a resized KB that can be directly plugged into the original VQ‑VAE encoder/decoder without modification.

The authors evaluate the approach on an image transmission task using the ImageNet dataset and a state‑of‑the‑art SimVQ VQ‑VAE backbone. They compare “Zero‑shot resizing” against a baseline where a dedicated KB is trained from scratch for each size. Results show that the zero‑shot method achieves SSIM scores virtually identical to the baseline across a wide range of K values, while requiring only a single training run. Notably, at very low rates (small K) the baseline suffers catastrophic degradation, whereas the proposed method degrades gracefully. Computationally, the method reduces GPU memory and training time by an order of magnitude because only the initial large KB needs to be learned; subsequent resizing is a lightweight graph operation.

The contribution is twofold: (i) a principled, hyperbolic‑geometry‑based technique for extracting a semantic hierarchy from a codebook, and (ii) a practical, zero‑shot resizing procedure that enables truly granular, on‑the‑fly rate adaptation in digital semantic communication. The paper also discusses limitations, such as the simplicity of leaf‑only pruning and the need for further studies on alternative pruning criteria, curvature tuning, and robustness under realistic noisy channel conditions. Overall, the work provides a compelling solution to the rate‑adaptation bottleneck in VQ‑VAE‑based semantic communication and opens avenues for more flexible, resource‑aware wireless systems.