Pith. sign in

REVIEW 2 major objections 5 minor 43 references

VQ4ALL: Efficient Neural Network Representation via a Universal Codebook

T0 review · 2 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read A single universal codebook compresses multiple network architectures past 16x while keeping accuracy near the floating-point original.

desk verdict A genuinely novel shared-codebook VQ method with strong low-bit accuracy, but the '>16×' compression headline only survives if you don't count the codebook—standard accounting puts a 2-bit ResNet-18 near 9.5×. read the letter →

arxiv 2412.06875 v1 pith:65KPRIL5 submitted 2024-12-09 cs.LG cs.AI

classification cs.LGcs.AI
keywords vectorquantizationuniversalcodebooknetworkcompressionkerneldensityestimationprogressiveconstructionlow-bitnetworksknowledgedistillationimageclassification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

VQ4ALL tries to establish that many different neural networks can be compressed from one shared, frozen codebook rather than from per-layer code tables. The paper builds a single codebook by kernel density estimation over weight sub-vectors sampled from several networks, then learns differentiable assignment ratios per network to reconstruct low-bit weights. If the claim is right, model fleets no longer multiply codebook storage: one static table can live in ROM, and switching tasks only swaps assignment indices. The reported support is compression above 16x with accuracies close to the floating-point baselines on ImageNet classification, COCO object detection, and text-to-image generation.

What carries the argument

The carrying mechanism is the universal codebook $C$ of $k$ codewords of dimension $d$, sampled from the kernel density estimate $f(w)$ of all networks' weight sub-vectors and then frozen. Each sub-vector keeps the $n$ nearest codewords as candidates, with softmax ratios $R$ initialized inversely to Euclidean distance, so the reconstructed weight $\widehat{W} = R C[A_c]$ is differentiable and gradients can move the ratios without touching the codebook. The Progressive Network Construction strategy then converts ratios above threshold $\alpha$ (0.9999) into one-hot optimal assignments gradually, which prevents the accuracy collapse observed when all assignments are switched at once.

What would settle it

Compute the deployed size of a single network as its assignment indices plus its share of the universal codebook, and compare with the floating-point model; for the paper's 2-bit ResNet-18 with a $2^{16}\times 8$ codebook, the ratio drops from 16x to about 9x, so a direct measurement of single-network or few-network storage would settle whether the amortization assumption holds.

Watch

Extended reading notes

Core claim

The central claim is that universal vector quantization, meaning one codebook shared by all networks, can replace per-layer codebooks without sacrificing accuracy. VQ4ALL extracts the codebook from a kernel density estimate of the concatenated weight sub-vectors of ResNet-18/50, MobileNet-V2, Mask R-CNN, and Stable Diffusion, freezes it, and represents each sub-vector by a shortlist of candidate codewords with softmax ratios. The ratios are trained against task loss, block-wise knowledge distillation, and a regularization term; a Progressive Network Construction strategy freezes each assignment to a one-hot choice once its ratio crosses a high threshold. With the codebook excluded from per-network size, the method reports 11x, 16x, and 32x compression at 3-bit, 2-bit, and 1-bit for classification, 26.3x for Mask R-CNN, and usable 2-3 bit Stable Diffusion, with accuracy near each floating-point original.

Load-bearing premise

The headline compression ratios count the universal codebook as free, assuming it is stored once in ROM and shared by so many networks that its size can be left out of each network's reported size.

Editorial extensions

If this is right

  • A deployed fleet of compressed networks can be served from one ROM-resident code table, so adding another network adds only assignment indices, not another codebook.
  • Memory access and silicon area for codebooks stop growing with the number of networks, since the universal table is loaded once rather than per layer or per model.
  • Extremely low-bit compression remains usable: at 1-bit, VQ4ALL reports 68.0% Top-1 on ResNet-18, 74.7% on ResNet-50, and 60.4% on MobileNet-V2 at 32x compression.
  • The same codebook transfers across task families, with 26.3x compression on Mask R-CNN detection and segmentation and 2-3 bit Stable Diffusion generation staying close to the floating-point FID and CLIP scores.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the per-network model size excludes the universal codebook, so the honest single-network ratio requires adding the codebook; the paper does not state how many networks are needed for the amortization to make that term negligible.
  • Editorial inference: if the codebook is truly reusable across architectures, the idea points toward a hardware primitive, one ROM code table plus per-network assignment storage, that would make the marginal storage for a new model very small.
  • Editorial inference: because only assignment learning is needed for a new network, the method could be tested as a continual-learning setting where new task models are added against a fixed codebook without retraining it.
  • Editorial inference: the same KDE-sampled universal codebook could be tried on other large-model families such as language models, where sub-vector statistics may also be shared across checkpoints.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper introduces VQ4ALL, a vector quantization method that uses a single shared 'universal codebook' to represent many low-bit neural networks. The codebook is initialized by kernel density estimation over weight sub-vectors sampled from the networks to be compressed, and each network is then constructed by learning differentiable soft assignments to candidate codewords, with a Progressive Network Construction (PNC) strategy that gradually freezes assignments. Experiments on ResNet-18/50, MobileNet-V2, Mask R-CNN, and Stable Diffusion report state-of-the-art accuracy at low bit-widths and claim compression rates exceeding 16x. The central claim depends on counting only the compressed index storage and excluding the universal codebook from the per-network model size, with the codebook assumed to be stored once in ROM and shared across an unspecified number of networks.

Significance. If the compression accounting is stated correctly, VQ4ALL is a meaningful step toward efficient multi-network deployment: a single code table in ROM promises large savings in memory access, codebook count, and silicon area, and the experimental results are strong across classification, detection, and generation. The paper deserves credit for a clean ablation study (Table 5, Figure 3) showing the contribution of PNC and candidate assignments, and for Table 6, which demonstrates that the codebook is not overly sensitive to the exact combination of source networks. The method is also practical: it runs on a single GPU and does not require the massive compute reported for some baselines. However, the headline compression ratios and the 'universal' characterization are currently overstated relative to the evidence, primarily because the codebook size is omitted from per-network size and because the codebook is always sampled from the very networks being evaluated.

major comments (2)
  1. [Abstract, Section 5.1, Table 3] The headline claim of 'compression rates exceeding 16x' is not supported under the standard definition of model size for a single network, because the universal codebook is excluded from the reported per-network size. The paper states in Section 5.1 that the universal codebook is 'stored as static code tables in the built-in ROM' and that Table 3 reports 'the average compression rate of compressed layers,' but it never adds the codebook back into the deployed model size or provides an amortization analysis over the number of networks sharing it. Concrete recalculation: for 2-bit ResNet-18 (k=2^16, d=8, 32-bit entries), the universal codebook is 2 MB; the compressed index stream for the compressed portion is about 2.9 MB, and the full-precision model is about 46.8 MB, so the true single-network compression ratio is about 9.5x, not 16x. At 1 bit (k=2^16, d=16), the codebook is 4 MB and the reported 32x ratio drops below 9x. The paper must disclose the codebook size in every compression-ratio report, state how many networks are required for the amortized codebook cost to become negligible, and present a comparison with DKM and other baselines using the same accounting for both methods.
  2. [Section 4.1, Section 5.4, Table 6] The claim that the codebook is 'universal' is only tested on networks whose weights were used to build it. The method description says the KDE samples from 'the networks mentioned above,' and the experimental setup in Section 5 states that sub-vectors are extracted from exactly the networks evaluated in Sections 5.1-5.3. Table 6 varies the combination of these same four networks, which shows robustness to the composition of the source set but does not test generalization to an architecture or dataset not seen during codebook construction. To support the 'universal' and 'applies to most deep neural networks' claims, the paper should evaluate a held-out network (e.g., ResNet-34, ViT-Tiny, or a different detection backbone) using a codebook built only from the other networks and report accuracy and compression under the same accounting.
minor comments (5)
  1. [Section 5.1] There is an internal contradiction about which layers are compressed in Table 3. The text says 'We do not compress the input layers ... The output layer is constructed using a small per-layer codebook derived from clustering its weights,' but later says 'Following the configuration of EWGS, none of the experiments in Table 3 compress the input and output layers of the network.' Please clarify which layers are actually included in the reported compression ratios.
  2. [Equation (11)] Equation (11) is typeset with unbalanced parentheses and an ambiguous denominator; it should be written as L_r = n * sum_{o,i/d,n} (r(1-r) / (o*i/d)) or equivalent, with a clear definition of the summation range.
  3. [Section 4.2] The update rule 'R ← R - u (∂/∂L, θ)' is not well-formed; it should explicitly state that R is updated with an optimizer u using the gradient ∂L/∂R and hyperparameters θ.
  4. [Table 1] In Table 1, the units of column 'C' are not specified (entries versus bytes), and the 'I/O' metric is not defined; please state how the number of codebook accesses is counted for the per-layer and universal variants.
  5. [Section 5.1] The text says that VQ4ALL 'continues to hold an accuracy close to 67%' when the compression ratio 'exceeds 40x,' but no table reports a 40x ratio; Figure 2 appears to show values in that range, so please label the figure axis clearly as applying to the compressed layers only and reconcile the numbers with the text.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported 16x/32x compression ratios are bit-width ratios (32/bit) computed over the index stream only; the fitted universal codebook is excluded, making the headline compression claim definitional rather than derived.

  1. self definitional [Section 3.1 (codebook storage formula); Section 5.1 and Table 3 (compression-ratio reporting)]
    "A can be stored using o×i/d × log2 k bits and C can be stored using k × d × 32 bits. ... We also report the average compression rate of compressed layers. ... VQ4ALL achieves optimal compression rates, as its universal codebook is shared with various low-bit networks and stored as static code tables in the built-in ROM."

    The headline compression ratios in Table 3 (16x at 2-bit, 32x at 1-bit) equal 32 divided by the bit width of the index stream; they count only the assignment storage and exclude the universal codebook. The paper's own formula fixes the codebook at k × d × 32 bits, so for a 2-bit ResNet-18 the 2 MB codebook (k=2^16, d=8) reduces the true single-network ratio from about 16x to about 9x. The codebook is itself a fitted parameter sampled from the weights of the compressed networks, and the number of networks needed to amortize it is never specified. Thus the 'compression rates exceeding 16x' claim is a definitional consequence of reporting per-layer index compression while treating the fitted codebook as free, not an empirically derived total-model compression ratio.

full rationale

The accuracy part of the paper is largely self-contained against external benchmarks: assignment ratios are fine-tuned with block-wise distillation on calibration data and evaluated on held-out ImageNet/COCO splits, and Table 6 shows a codebook built from ResNet-18 alone transfers to ResNet-50, providing independent evidence that the universal codebook is not merely memorizing the evaluated networks. The citation to the authors' VQ4DiT is a minor related-work pointer and is not load-bearing; no uniqueness theorem or ansatz is smuggled in via self-citation. The main circularity is confined to the compression-ratio accounting, which is one of the two headline claims: the 16x/32x numbers are the bit-width ratios of the index stream, with the fitted universal codebook excluded by construction. Because the central quantitative compression claim reduces to that definitional choice, the score is 6; the accuracy results remain empirically grounded and independent.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim depends on several hand-chosen hyperparameters (bandwidth, codebook shape, candidate count, PNC threshold) and on domain assumptions about KDE representativeness, block-wise distillation transfer, and ratio convergence. No new physical or conceptual entities are introduced.

free parameters (5)
  • KDE bandwidth h = 0.01
    Controls smoothing of the kernel density estimate used to sample the universal codebook; set by hand in Section 5 hyperparameters.
  • Universal codebook size (k, d) = 3-bit: 4096x4; 2-bit: 65536x8; 1-bit: 65536x16; 0.5-bit: 65536x32
    Chosen by hand to match target bit-widths; directly determines both compression ratio and codebook storage cost (Section 5 hyperparameters).
  • Candidate assignments n = 64
    Number of nearest codewords considered per sub-vector; set to 64 in all experiments and ablated in Table 5.
  • PNC threshold alpha = 0.9999
    Ratio threshold for freezing one-hot assignments in Progressive Network Construction; searched in supplementary Figure 4.
  • Ratio learning rate = 3e-1 (Adamax)
    Learning rate for assignment ratios; chosen without per-task tuning, as stated in Section 5.
assumptions (4)
  • domain assumption Kernel density estimation of weight sub-vectors yields a sampling distribution covering each network's weight distribution well enough for low-bit reconstruction.
    The universal codebook is sampled from a KDE fit to a concatenation of sub-vectors from the target networks; this assumes the sample is representative of each individual network's layers (Section 4.1).
  • domain assumption Block-wise knowledge distillation from the floating-point network to the low-bit network preserves final task accuracy.
    The objective includes L_kd, which aligns intermediate block features; the method assumes this alignment transfers to task performance (Section 4.2).
  • ad hoc to paper Softmax ratios trained with the L_r regularizer converge toward one-hot values, and freezing them at alpha=0.9999 avoids accuracy collapse.
    Progressive Network Construction is an empirical strategy; the paper observes convergence in Figure 3 but provides no theoretical guarantee that ratios will concentrate near 1 for arbitrary layers or networks (Section 4.3).
  • domain assumption The calibration dataset used for fine-tuning assignments is representative of the validation distribution.
    All experiments use the original training and validation data (ImageNet, COCO) for calibration and evaluation; the paper does not test with disjoint calibration data (Section 5).

how reviews work

0 comments
Cite this review

Pith. "Pith review of VQ4ALL: Efficient Neural Network Representation via a Universal Codebook." pith.science (2026). https://pith.science/paper/65KPRIL5

@misc{pith2026241206875,
  author       = {Pith},
  title        = {Pith review of: VQ4ALL: Efficient Neural Network Representation via a Universal Codebook},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/65KPRIL5}},
  note         = {Machine review of arXiv:2412.06875}
}
abstract

The rapid growth of the big neural network models puts forward new requirements for lightweight network representation methods. The traditional methods based on model compression have achieved great success, especially VQ technology which realizes the high compression ratio of models by sharing code words. However, because each layer of the network needs to build a code table, the traditional top-down compression technology lacks attention to the underlying commonalities, resulting in limited compression rate and frequent memory access. In this paper, we propose a bottom-up method to share the universal codebook among multiple neural networks, which not only effectively reduces the number of codebooks but also further reduces the memory access and chip area by storing static code tables in the built-in ROM. Specifically, we introduce VQ4ALL, a VQ-based method that utilizes codewords to enable the construction of various neural networks and achieve efficient representations. The core idea of our method is to adopt a kernel density estimation approach to extract a universal codebook and then progressively construct different low-bit networks by updating differentiable assignments. Experimental results demonstrate that VQ4ALL achieves compression rates exceeding 16 $\times$ while preserving high accuracy across multiple network architectures, highlighting its effectiveness and versatility.

Figures

Figures reproduced from arXiv: 2412.06875 by the authors.

Figure 1
Figure 1. The bottom-up pipeline of VQ4ALL. The Universal Codebook is randomly sampled from the kernel density estimation of the floating-point sub-vectors of several networks. Based on the codebook, differentiable candidate assignments are assigned to each type of network, all initialized with the same ratios. The progressive network construction strategy then calibrates the ratios and sets candidate assignments with high ra… view at source ↗
Figure 2
Figure 2. Compression results for ResNet-18 and ResNet-50. We compare the trade-off between accuracy and compression ratio, using pre-trained models from the PyTorch zoo as baselines. Overall, our method demonstrates superior accuracy compared to previous approaches. Size Ratio APbb APbb 50 APbb 75 APmk APmk 50 APmk 75 Mask-RCNN R-50 FPN (uncompressed) 169.40 MB 1.0× 37.9 59.2 41.1 34.6 56.0 36.8 RetinaNet (uncompressed) 145.… view at source ↗
Figure 3
Figure 3. Ablation results on Progressive Network Construc￾tion (PNC) Strategy with the same 2-bit ResNet-18 compression configuration. Up: we compare the accuracy of VQ4ALL with and without PNC at each epoch. Once the VQ4ALL pipeline is complete, we convert the candidate assignments with the largest ratios in VQ4ALL (no PNC) to optimal assignments, resulting in a drop in accuracy to 57.77%. Down: in the distribution of the l… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: 2-bit ResNet-18/50 compression with varying ratio [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Optimal assignment distribution of various low-bit Net [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: 3-bit Stable Diffusion v1-4 image generation using COCO prompts. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: 2-bit Stable Diffusion v1-4 image generation using COCO prompts. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 23 canonical work pages

  1. [1]

    Metaquant: Learning to quantize by learning to penetrate non-differentiable quantization

    Shangyu Chen, Wenya Wang, and Sinno Jialin Pan. Metaquant: Learning to quantize by learning to penetrate non-differentiable quantization. In Advances in Neural In- formation Processing Systems, 2019. 6

  2. [2]

    Dkm: Differentiable k-means cluster- ing layer for neural network compression

    Minsik Cho, Keivan A Vahid, Saurabh Adya, and Mo- hammad Rastegari. Dkm: Differentiable k-means cluster- ing layer for neural network compression. arXiv preprint arXiv:2108.12659, 2021. 2, 6

  3. [3]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 5

  4. [4]

    Vq4dit: Efficient post-training vec- tor quantization for diffusion transformers

    Juncan Deng, Shuaiting Li, Zeyu Wang, Hong Gu, Kedong Xu, and Kejie Huang. Vq4dit: Efficient post-training vec- tor quantization for diffusion transformers. arXiv preprint arXiv:2408.17131, 2024. 2

  5. [5]

    Hawq-v2: Hessian aware trace-weighted quantization of neural networks

    Zhen Dong, Zhewei Yao, Daiyaan Arfeen, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Hawq-v2: Hessian aware trace-weighted quantization of neural networks. Ad- vances in neural information processing systems, 33:18518– 18529, 2020. 7

  6. [6]

    Training with quantization noise for extreme model com- pression

    Angela Fan, Pierre Stock, Benjamin Graham, Edouard Grave, R ´emi Gribonval, Herv ´e J ´egou, and Armand Joulin. Training with quantization noise for extreme model com- pression. In International Conference on Learning Repre- sentations, 2021. 2

  7. [7]

    Deep neural network com- pression by in-parallel pruning-quantization

    Tung Frederick and Mori Greg. Deep neural network com- pression by in-parallel pruning-quantization. IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 2018. 6

  8. [8]

    Compressing deep convolutional networks using vector quantization

    Yunchao Gong, Liu Liu, Ming Yang, and Lubomir Bour- dev. Compressing deep convolutional networks using vector quantization. arXiv preprint arXiv:1412.6115, 2014. 2

Show all 43 references
  1. [9]

    Song Han, Huizi Mao, and William J. Dally. Deep com- pression: Compressing deep neural network with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016. 2, 6

  2. [10]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 5

  3. [11]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Gir- shick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017. 5

  4. [12]

    Clipscore: A reference-free evaluation met- ric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation met- ric for image captioning. arXiv preprint arXiv:2104.08718,

  5. [13]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. Advances in neural information processing systems , 30, 2017. 7

  6. [14]

    B. Ham J. Lee, D. Kim. Network quantization with element- wise gradient scaling. In Proceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition, 2021. 2, 7

  7. [15]

    Categorical reparameterization with gumbel-softmax

    Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 6

  8. [16]

    Adam: A method for stochastic opti- mization

    Diederik P Kingma. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014. 6

  9. [17]

    Fully quantized network for object detection

    Rundong Li, Yan Wang, Feng Liang, Hongwei Qin, Jun- jie Yan, and Rui Fan. Fully quantized network for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2810–2819,

  10. [18]

    Q-diffusion: Quantizing diffusion models

    Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 17535–17545, 2023. 8

  11. [19]

    Focal loss for dense object detection

    T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 7

  12. [20]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceeding...

  13. [21]

    Towards accurate binary convolutional neural network

    Xiaofan Lin, Cong Zhao, and Wei Pan. Towards accurate binary convolutional neural network. Advances in neural in- formation processing systems, 30, 2017. 6

  14. [22]

    Sgdr: Stochas- tic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochas- tic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6

  15. [23]

    Permute, quantize, and fine-tune: Efficient compression of neural networks

    Julieta Martinez, Jashan Shewakramani, Ting Wei Liu, Ioan Andrei B ˆarsan, Wenyuan Zeng, and Raquel Urtasun. Permute, quantize, and fine-tune: Efficient compression of neural networks. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pag...

  16. [24]

    Profit: A novel training method for sub-4-bit mobilenet models

    Eunhyeok Park and Sungjoo Yoo. Profit: A novel training method for sub-4-bit mobilenet models. In European Con- ference on Computer Vision, 2020. 2 9

  17. [25]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

  18. [26]

    Xnor-net: Imagenet classification using bi- nary convolutional neural networks

    Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using bi- nary convolutional neural networks. In European conference on computer vision, pages 525–542. Springer, 2016. 6

  19. [27]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 5

  20. [28]

    Improved techniques for training gans

    Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 7

  21. [29]

    Mobilenetv2: Inverted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh- moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition ,

  22. [30]

    Learning dis- crete weights using the local reparameterization trick

    Oran Shayer, Dan Levi, and Ethan Fetaya. Learning dis- crete weights using the local reparameterization trick. arXiv preprint arXiv:1710.07739, 2017. 6

  23. [31]

    Cluster- ing convolutional kernels to compress deep neural networks

    Sanghyun Son, Seungjun Nah, and Kyoung Mu Lee. Cluster- ing convolutional kernels to compress deep neural networks. In Proceedings of the European conference on computer vi- sion (ECCV), pages 216–232, 2018. 2

  24. [32]

    And the bit goes down: Re- visiting the quantization of neural networks

    Pierre Stock, Armand Joulin, R ´emi Gribonval, Benjamin Graham, and Herv ´e J ´egou. And the bit goes down: Re- visiting the quantization of neural networks. arXiv preprint arXiv:1907.05686, 2019. 2, 6

  25. [33]

    Post-training quan- tization with progressive calibration and activation relax- ing for text-to-image diffusion models

    Siao Tang, Xin Wang, Hong Chen, Chaoyu Guan, Zewen Wu, Yansong Tang, and Wenwu Zhu. Post-training quan- tization with progressive calibration and activation relax- ing for text-to-image diffusion models. arXiv preprint arXiv:2311.06322, 2023. 8

  26. [34]

    Haq: Hardware-aware automated quantization with mixed precision

    Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. Haq: Hardware-aware automated quantization with mixed precision. In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition, 2019. 2, 6

  27. [35]

    Quantized convolutional neural networks for mobile devices

    Jiaxiang Wu, Cong Leng, Yuhang Wang, Qinghao Hu, and Jian Cheng. Quantized convolutional neural networks for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4820–4828,

  28. [36]

    Gobo: Quantizing attention-based nlp models for low latency and energy efficient inference

    Ali Hadi Zadeh, Isak Edo, Omar Mohamed Awad, and An- dreas Moshovos. Gobo: Quantizing attention-based nlp models for low latency and energy efficient inference. In 2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture, pages 811–824. IEEE, 2020. 2

  29. [37]

    Trained ternary quantization

    Chenzhuo Zhu, Song Han, Huizi Mao, and William J Dally. Trained ternary quantization. arXiv preprint arXiv:1612.01064, 2016. 5 10 VQ4ALL: Efficient Neural Network Representation via a Universal Codebook Supplementary Material 0.80 0.85 0.90 0.95 0.9999 68 70 72 74 76T op-1 Acc...

  30. [38]

    Ratio Threshold α Searching As shown in Figure 4, we evaluate the impact of different ra- tio threshold α values in the Progressive Network Construc- tion Strategy on 2-bit ResNet-18/50. The results demon- strate that smaller α values lead to lower network accu- racy, indicati...

  31. [39]

    As shown in Figure 5, each type of low-bit network is evenly composed of differ- ent codewords of the same universal codebook

    Optimal Assignment Distribution To verify whether the codewords of the universal codebook are fully utilized, we analyzed the optimal assignment dis- tribution across different networks. As shown in Figure 5, each type of low-bit network is evenly composed of differ- ent codew...

  32. [40]

    We then evaluate the impact of these codebooks on network performance

    Design of the Universal Codebook As shown in Table 6, we adopt Kernel Density Estimation (KDE) to fit the combinations of weights from different net- works and randomly sample to generate various universal codebooks. We then evaluate the impact of these codebooks on network pe...

  33. [41]

    Random candidate assignments yield the poorest performance, with the accuracy of 2-bit ResNet-18 and 2-bit ResNet-50 dropping to only 39.97% and 44.36%, respec- tively

    Configuration of Candidate Assignments As shown in Table 7, we evaluated the impact of different configurations of candidate assignments on network perfor- mance. Random candidate assignments yield the poorest performance, with the accuracy of 2-bit ResNet-18 and 2-bit ResNet-...

  34. [42]

    For ResNet-18/50, the primary blocks are ’Ba- sicBlock’ and ’Bottleneck’

    Settings of Block-Wise Knowledge Distilla- tion Section 4.2 introduces the objective function of block-wise knowledge distillation, where different networks are di- vided into their respective primary blocks based on their structure. For ResNet-18/50, the primary blocks are ’B...

  35. [43]

    Compared to other state-of-the-art uniform quan- tization methods, the images generated by VQ4ALL are more similar to those produced by the floating-point net- work

    Comparison of Generated Images In Figure 6 and Figure 7, we present the images generated by the low-bit Stable Diffusion v1-4 network on COCO prompts. Compared to other state-of-the-art uniform quan- tization methods, the images generated by VQ4ALL are more similar to those pr...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.