Skip to main content
Search

# Vector Quantization

7 related articles

Subscribe to updates
  1. arXiv

    Detector data squeezed into discrete tokens: tracks and topoclusters hold 1% jet resolution at 4–11x

    High-energy-physics detector data can be compressed into discrete tokens, with track and topocluster tokenizers holding jet resolution below 1% at roughly 4–11x compression, while calorimeter cells stay slightly above that target.

  2. arXiv

    Skeleton action recognition gets a quantization plug-in: white-box attack success drops from near 100% to single digits, accuracy stays comparable

    Mainstream contrastive models for self-supervised skeleton-based action recognition had almost no defense against adversarial attacks, where tiny perturbations to skeleton data could sharply degrade recognition. QCLR adds quantization as a plug-in module at two levels of the encoding pipeline, mapping continuous inputs and features to discrete codebooks so small perturbations are absorbed and gradient propagation is broken; integrated into SkeletonCLR, AimCLR and ActCLR on NTU-60, NTU-120 and PKU-MMD, it cuts PGD and SMART attack success rates to single digits while keeping classification accuracy comparable to the original models.

  3. arXiv

    Polytopal Neural Networks constrain latent representations to learned polytopes, retaining 96.9% of unconstrained accuracy on ImageNet-1k

    The authors propose Polytopal Neural Networks (PNNs), which constrain each layer's latent representation during training to a convex combination of learned archetypes so that the explanatory coordinates are the same coordinates the next layer receives; across image classification and reconstruction benchmarks, PNNs approach unconstrained networks with small accuracy loss, outperform VQ-VAE and Dirichlet VAE in unsupervised reconstruction, and offer a VQ training route without EMA updates, commitment loss, or straight-through estimator.

  4. arXiv

    Learning Latent Protein Languages: PLLM Raises Compute-Scaling Exponent to 0.038 and SLL Cuts Validation Perplexity by 34%

    This work introduces two learned discrete latent protein languages: PLL, a 4,096-state contextual alphabet built on a frozen ESM-2 encoder on the sequence side, and SLL, which adapts GCP-VQVAE Lite with auxiliary sequence and confidence supervision while remaining decodable to backbone coordinates; under matched downstream training, PLLM has a fitted compute-scaling exponent of 0.038 versus 0.020 for the amino acid autoregressive model and reduces the fraction of samples below a 1.5-bit residue-composition entropy threshold by 54%, while replacing the original tokenizer with SLL reduces best validation perplexity by 34% for sequence-to-structure prediction.

  5. arXiv

    AIR-MoE routes with vector-quantized two-stage inverted indexing, improving the perplexity-FLOPs trade-off at over a million tiny experts

    The work introduces AIR-MoE, an inverted-index-inspired two-stage MoE routing architecture based on vector quantization: a first stage assigns tokens to VQ codewords to build a candidate expert set, and a second stage computes exact routing scores restricted to that shortlist, approximating true top-K routing while avoiding full expert scoring and imposing no structural constraints on expert parameters; the authors also provide a lower bound on the mass recall achieved by AIR-MoE, and empirically improve the perplexity-FLOPs trade-off over existing MoE routers in regimes with up to more than a million tiny experts while achieving consistent perplexity improvements over the best existing baseline.

  6. arXiv

    Across 25 models, a shared L0 attention routing circuit marks VQ-tokenized VLMs, and ablating it alone cuts open-ended object hallucination by 31% relative

    Using activation patching across twenty-five models spanning eight LLM families, the work identifies an early-layer (L0) attention routing circuit shared by VQ-tokenized vision-language models, proposes a three-gate diagnostic that isolates ten models carrying it and rejects the other fifteen, shows via a single-variable architectural swap (LLaVA-1.6 CLIP+MLP to VQ+Linear) that vector quantization is the source of the pathological signal, and finds that only L0 ablation reduces object hallucination in open-ended generation (CHAIR_i down 31% relative) while tuned DoLA and VCD do not.

  7. arXiv

    StableVQ: Turning Vector-Quantized Tokenizer Training Stability from a Fragile Outcome into a Guaranteed Property

    The work proposes StableVQ, three parameter-free components — Dynamic STE, Region VQ Loss, and Decoupled Schedule — that disentangle Encoder–Decoder and Codebook training so each module can fulfill its own responsibility independently, consistently improving training stability, codebook utilization, and reconstruction quality on ImageNet across diverse codebook sizes and initialization settings.

You’re all caught up