vector quantization

Existing global shared codebook approaches neglect intra-face semantic correlations and token-level semantic disparities, leading to suboptimal reconstruction quality and face recognition performance at ultra-low bitrates (e.g., 0.05 bpp). To address this, we propose a switchable token-specific codebook quantization framework: first, codebooks are learned independently per semantic category; then, each visual token is dynamically assigned its most suitable dedicated codebook, enabling fine-grained, low-distortion quantization. Our method is the first to jointly couple token-level codebook selection with category-aware grouping—reducing individual codebook size while enhancing representational diversity and quantization fidelity. Experiments demonstrate that reconstructed face images achieve a mean recognition accuracy of 93.51% at 0.05 bpp, significantly outperforming global codebook baselines. This work establishes a novel paradigm for codebook-driven face compression models.

In whole-brain MRI segmentation with a large number of classes (108), one-hot encoding incurs prohibitive computational and memory overhead. Method: This work proposes a compact label representation paradigm based on binary encoding—introducing, for the first time in medical image segmentation, logarithmic-complexity coding schemes (e.g., Error-Correcting Output Codes), label embedding trees, and soft/hard decoding strategies. It incorporates class-to-codeword optimization, class-weighted loss adaptation, and error-correction mechanisms to reduce model parameters and GPU memory consumption. Contribution/Results: Although current binary-encoded models achieve lower Dice scores (DSC = 39.3–73.8) than the one-hot baseline (DSC = 82.4), they reveal a fundamental trade-off among encoding robustness, decoding accuracy, and semantic consistency. The approach demonstrates feasibility and scalability for large-scale, fine-grained medical segmentation, offering a novel, resource-efficient pathway beyond conventional label representations.

In goal-oriented communication, there is a need for direct learning in the compressed domain without full image reconstruction. Method: This paper proposes a structure-preserving compression coding paradigm that replaces conventional entropy coding (e.g., Huffman or arithmetic coding) with low-density parity-check (LDPC) codes—deployed at the entropy coding layer—for end-to-end compressed-domain classification. Leveraging the sparse graph structure of LDPC codes, semantic correlations are inherently preserved. A gated recurrent unit (GRU)-based compressed-domain classifier is designed and jointly optimized with the encoder via end-to-end training. Contribution/Results: Experiments on MNIST, Fashion-MNIST, and CIFAR-10 demonstrate superior classification accuracy over traditional entropy coding baselines, with an order-of-magnitude reduction in model parameters, significantly lowering computational overhead and deployment complexity.

To address low quantization efficiency, poor robustness, and semantic distortion induced by channel noise in digital semantic communication, this paper proposes an information-theoretic, learnable codebook design framework. We first establish a theoretical equivalence between semantic synonym mapping and Voronoi partitioning, then formulate an end-to-end jointly optimized objective comprising a semantic-maximizing entropy-regularized quantization loss and a channel-aware semantic distortion loss. The method integrates mutual information maximization, Voronoi-based quantization modeling, and channel distortion characterization. Evaluated on image reconstruction, the proposed approach achieves a 24.1% PSNR gain and a 46.5% improvement in LPIPS perceptual similarity at 10 dB SNR, significantly mitigating semantic distortion. This work introduces a novel paradigm for efficient and reliable semantic-driven transmission.

Training and deploying large deep learning models under resource constraints faces significant challenges in computational cost, memory footprint, and communication overhead. This paper proposes Coding Deep Learning (CDL), the first framework to embed information-theoretic entropy constraints directly into the deep learning kernel, enabling joint differentiable probabilistic quantization of weights and activations while strictly enforcing entropy bounds throughout end-to-end training—ensuring intrinsic compressibility. We further introduce R-CDL, a variant optimizing the accuracy–compression trade-off. CDL integrates quantization-aware training, soft differentiable approximations, and parallelized compressed communication. Experiments demonstrate that CDL comprehensively outperforms state-of-the-art methods on DNN compression: it substantially reduces training FLOPs, significantly lowers inference latency and model size, and enables plug-and-play deployment with ultra-low-bitwidth (≤4-bit) quantization.

This work addresses the challenge of maintaining perceptual quality consistency in learned video codecs when deployed across varying spatial resolutions, a scenario that typically necessitates retraining or rate-distortion parameter tuning. Building upon the MS-VQ-VAE framework, the study systematically investigates the impact of codebook capacity and spatial resolution on perceptual quality using the UCF101 dataset. The findings reveal that codebook capacity exerts an influence approximately ten times greater than that of resolution, with higher resolutions yielding superior entropy efficiency. These insights offer a novel perspective for designing discrete tokenizers in multi-resolution video compression and generative models. Experimental results demonstrate that the proposed method achieves LPIPS scores surpassing H.264 by 25–52% at 128×128 resolution and outperforming H.265 by 21–37% at 256×256, all while operating at comparable or lower bitrates.

This work addresses the practical limitations of vector quantization for compressing large language model weights, which typically relies on input-specific statistics and thus lacks generalizability. The study investigates whether a universal codebook can achieve near-optimal quantization performance without depending on any particular input distribution. By integrating tools from information theory, PCA alignment, water-filling allocation, and high-dimensional spherical covering, the authors theoretically establish—for the first time—that such a universal codebook exists, whose rate-distortion performance is at most 0.11 bits per dimension worse than that of an input-optimized water-filling codebook. This result demonstrates the near-optimality of universal quantization schemes and provides a theoretical foundation for low-precision model storage, although explicit constructions of such codebooks remain an open challenge.

This work addresses the dimensional collapse of latent representations in Vector Quantized Variational Autoencoders (VQ-VAEs), where vector quantization often confines the latent space to a low-dimensional subspace, degrading both reconstruction fidelity and perceptual quality. The paper provides the first mechanistic explanation of this phenomenon through the lenses of rate–distortion theory and sequential learning dynamics. To mitigate this issue without altering the model architecture, the authors propose an AE Warm-Up strategy: pretraining the encoder–decoder as a continuous autoencoder prior to introducing quantization, thereby preserving high-dimensional latent structure. Evaluated on both image (VQGAN) and audio (WavTokenizer) tasks, this approach substantially increases the effective latent dimensionality—from 3–5 to 17–19—while reducing rFID by 17–35% and improving PESQ by 11–14%, all without additional training cost.

Clustering high-dimensional discrete data is often hindered by high computational cost, sensitivity to sparsity, and limited methodological applicability. This work proposes a deterministic dimensionality reduction framework that compresses binary, categorical, or count-based high-dimensional discrete data into low-dimensional continuous representations via weighted positional encoding. The resulting mapping is injective, preserving the discriminative structure of the original data; under mild conditions, the compressed variables approximately follow a Gaussian distribution while maintaining inter-cluster distances, thereby ensuring identifiable clustering structures. Empirical evaluations on real-world datasets—including infant names and microbiome profiles—demonstrate that the method achieves high clustering accuracy and substantially outperforms mainstream dimensionality reduction techniques in computational efficiency, offering both strong theoretical guarantees and practical utility.

This work addresses a fundamental reliability challenge in single-shot joint source-channel coding: ensuring that at least one decoder can reconstruct the source within a prescribed maximum distortion constraint over independent channels. To this end, the paper introduces, for the first time, a codebook diversity gain by designing disjoint codebooks for different decoders and synergistically combining it with channel diversity to enhance system robustness. The authors develop an extended multi-decoder framework grounded in the Poisson matching lemma and propose a hybrid encoding and grouping strategy that achieves an optimal trade-off between codebook diversity and channel diversity. Experimental results on binary symmetric channels demonstrate that the proposed hybrid scheme significantly outperforms baseline approaches employing either fully shared or entirely disjoint codebooks.