Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A searched low-rank correction plus a focused Softmax quantizer keeps low-bit vision transformers accurate.

desk verdict The method has good ideas, but the comparison tables look broken—several baselines are near chance level and there is no code or seeds to check. read the letter →

arxiv 2502.04628 v1 pith:OT6HH7TF submitted 2025-02-07 cs.CV

Runqing Jiang , Ye Zhang , Longguang Wang , Pengpeng Yu , Yulan Guo This is my paper · ORCID
classification cs.CV
keywords post-trainingquantizationvisiontransformerslow-bitlow-rankcompensationdifferentiablearchitecturesearchdynamicfocusingquantizersoftmaxactivationmodelcompression
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

This paper tries to establish that vision transformers can be post-training quantized down to 3- and 4-bit weights and activations without collapsing, by correcting the two losses that prior methods underweight: information destroyed when weights are rounded, and resolution wasted when post-Softmax attention maps are quantized with logarithmic transforms. The proposed AIQViT attaches a learnable low-rank correction to each linear layer, with each layer's rank chosen by a differentiable architecture-style search on the calibration set, and replaces log2 quantizers with a dynamic focusing quantizer that learns which interval of Softmax outputs deserves uniform quantization. On ImageNet, these components lift W3/A3 DeiT-T from the 14-16% range of prior methods to 38.51%, and W4/A4 DeiT-B to 79.19% against an 81.80% full-precision baseline; the same recipe transfers to COCO detection and segmentation and to point-cloud tasks. If the claim holds, low-bit ViTs become practical on memory-constrained devices without retraining on full datasets.

What carries the argument

The load-bearing object is the architecture-informed low-rank compensation: every quantized linear layer computes $h = \mathrm{Quant}_{\mathrm{U}}(x)\,\mathrm{Quant}_{\mathrm{U}}(W_0 + BA)$, with $A\in\mathbb{R}^{r\times m}$ and $B\in\mathbb{R}^{d\times r}$ trained against the full-precision block output, and with the per-layer rank chosen by differentiable architecture search from $\{10,20,50,100,150\}$. Its partner is the dynamic focusing quantizer, which learns endpoints $b_1,b_2$ and applies uniform quantization only on $[b_1,b_2]$, mapping the tails to the extremes. These two mechanisms carry the argument: the ablations show that removing either one costs 10-15 accuracy points at 3-bit, and the rank search alone is worth several points over any fixed rank tested.

What would settle it

Take two disjoint 1,024-sample calibration splits from ImageNet. Run the full AIQViT rank search and interval learning on the first split, then evaluate on the second split or on an out-of-domain set. If the automatically searched ranks stop beating the best fixed rank, such as the reported 38.51 versus 36.87 for DeiT-T W3/A3, or if the gap shrinks when the calibration split is resampled, then the architecture-search component is fitting calibration noise rather than transferring structure.

Watch

Extended reading notes

Core claim

AIQViT's central claim is that the two failure modes of low-bit ViT quantization are separable, and each can be fixed by a calibration-driven mechanism. The first is weight-quantization error, treated not as a quantizer-design problem but as a low-rank additive perturbation: for each fully connected layer the pretrained weight $W_0$ is frozen and quantized, while a learnable low-rank product $BA$ is added and optimized by reconstructing the block output, $\mathcal{L}_{\mathrm{rec}}(D_t^{(l)},\Theta^{(l)}) = \mathbb{E}_{x\in D_t^{(l)}} \| B_q^{(l)}(x) - B^{(l)}(x) \|_F$. The rank $r$ is not hand-tuned: a differentiable architecture search over $\{10,20,50,100,150\}$ selects the most useful rank per layer through a bilevel objective on a split of the calibration set. The second failure mode is post-Softmax activation quantization. The paper argues that the standard log2 quantizer spends its levels on values near zero that carry redundant attention information, so AIQViT instead learns interval endpoints $b_1$ and $b_2$, clips values outside $[b_1,b_2]$ to $0$ and $2^k-1$, and applies uniform quantization inside the selected interval. A curriculum-learning schedule that feeds easy calibration samples before hard ones stabilizes the reconstruction. Together these mechanisms produce the reported gains across Tables 1-6 and across image and point-cloud tasks.

Load-bearing premise

The load-bearing premise is that choosing each layer's rank by a search run on the same 1,024-sample calibration set used for fitting finds corrections that help on new test data; if the search merely memorizes calibration samples, the reported gains disappear outside that set.

Editorial extensions

If this is right

  • Memory and compute for ViTs can drop by roughly 4-8x in weight and activation bit-width while staying close to full precision, without training on the full dataset.
  • Because the low-rank corrections are merged into the quantized weights after calibration, inference needs no extra parameters or specialized logarithmic units; it runs with standard uniform kernels.
  • Per-layer rank should be treated as a searched hyperparameter: fixed ranks lose to automatically chosen ranks in the paper's ablations, for example DeiT-T W3/A3 at 38.51 versus 36.87 with r=100.
  • The recipe transfers across modalities: the same candidate rank set and reconstruction schedule work for image classification, detection and segmentation backbones, and point-cloud transformers.

Reading between the lines

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

  • Implicit but untested: if the searched ranks reflect genuine layer-wise structure, the chosen rank distribution should be stable across different calibration draws; the paper does not report this stability, and testing it would separate signal from calibration noise.
  • The DFQ principle generalizes: any long-tailed or sharply peaked activation distribution may be better served by a learned informative interval with uniform quantization than by a fixed nonlinear transform over the whole support, so the idea could extend beyond Softmax to activations like post-GELU features.
  • Since calibration uses only 1,024 or 256 samples, the low-rank corrections risk overfitting to the calibration domain; a natural extension is measuring AIQViT's accuracy under domain shift, such as quantizing on ImageNet and testing on natural distribution shifts.
  • The curriculum schedule's benefit being largest at 3-bit hints that sample ordering can substitute for extra calibration data in ultra-low-bit regimes; a direct comparison against simply doubling the calibration set would quantify that trade.
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

4 major / 6 minor

Summary. The paper proposes AIQViT, a post-training quantization method for vision transformers that combines (i) architecture-informed low-rank compensation matrices per linear layer, with ranks selected by a differentiable architecture search, and (ii) a dynamic focusing quantizer (DFQ) that learns the interval of post-Softmax activations to quantize uniformly. A curriculum-learning schedule orders calibration samples by reconstruction loss. The method is evaluated on ImageNet classification, COCO detection/segmentation, ModelNet40 classification, and ShapeNetPart segmentation, with ablations on DeiT-T and DeiT-S. The central claim is that AIQViT outperforms state-of-the-art PTQ methods on these tasks, especially at W3/A3 and W4/A4.

Significance. If the reported results are correct, AIQViT would be a practically useful low-bit PTQ method that works across image and point-cloud transformers, and the architecture-informed rank search is a conceptually interesting departure from fixed-rank LoRA-style compensation. The internal ablations are consistent: each of AILoC, DFQ, and curriculum learning contributes, and auto-selected ranks outperform fixed ranks. The evaluation across five vision tasks is a strength. However, the central SOTA claim rests on comparison tables whose baseline numbers are implausibly low and inconsistent with published results for the same methods. Because no code, seeds, or error bars are provided, the experimental evidence is not yet sufficient to establish the claimed superiority.

major comments (4)
  1. [Table 1] Several baseline entries are at chance level and inconsistent with published results for the same methods: RepQ-ViT W3/A3 reports 0.44% top-1 for ViT-S and 0.17% for ViT-B; QDrop W4/A4 DeiT-T reports 31.65% against a 72.21% full-precision model; BRECQ W3/A3 DeiT-S reports 8.26%. RepQ-ViT and QDrop are established low-bit ViT PTQ methods, so these numbers strongly suggest an incompatible calibration or evaluation protocol for the baselines rather than genuine performance. Since the paper's central superiority claim is derived almost entirely from Tables 1-4, this comparison is load-bearing and must be corrected. Please rerun every baseline with the official released code, report the exact quantization settings (softmax handling, scale reparameterization, calibration split), and provide seeds and error bars.
  2. [Table 2] The detection/segmentation baselines are similarly implausible. PTQ4ViT reports 6.5 box AP / 6.6 mask AP for Mask R-CNN with Swin-S at W6/A6, and 14.7/13.5 for Cascade Mask R-CNN with Swin-T at W4/A4. These values are far below the same method's W4/A4 Swin-S row (26.7/26.6) and below the level expected from a functional detection pipeline. Such numbers indicate that the PTQ4ViT baseline was not correctly integrated into the detection/segmentation framework, which inflates AIQViT's reported margin. The COCO experiments should be redone with verified baseline implementations, and the full corrected table should be reported.
  3. [Sec. 'The Choice of Rank r', Eqs. (13)-(14)] The rank search is a bilevel problem solved with an approximate DARTS gradient, where the validation split is drawn from the same 1,024-sample calibration set used for reconstruction. The paper provides no evidence that the searched ranks transfer beyond this calibration split; the entire AILoC gain in Table 5 (e.g., +15.31% at W3/A3 DeiT-T) is contingent on that transfer. Please report per-layer searched ranks for each model, compare auto-selected ranks against fixed ranks on a true held-out validation split, or repeat the search with different random calibration subsets to show stability.
  4. [Experiments] No code, seeds, or error bars are provided for any experiment. Given the small calibration sets (256-1,024 samples), the stochastic optimization (drop-path, random batches, curriculum ordering), and the fact that several reported margins are small (e.g., 0.7% in W6/A6 ImageNet rows), single-run numbers without variance do not establish the claimed improvements. Please provide an artifact/code link and report mean and standard deviation over multiple runs for both AIQViT and the baselines.
minor comments (6)
  1. [Eq. (9)] The notation 'Quant-U(x)Quant-U(W0 + BA)' is ambiguous: it is unclear whether BA is merged into W0 before weight quantization, or whether the low-rank compensation is kept in floating point and added to the quantized output. This affects both the derivation and the claimed hardware compatibility of the method.
  2. [Eqs. (11)-(12)] The symbol alpha is used both for the raw architecture parameter and for its softmax-normalized value; please use a distinct notation (e.g., alpha vs. alpha-bar) to avoid confusion.
  3. [Eq. (16)] The definition of D_t as an argmin over subsets of size lambda(t)|D| is not directly computable; please specify that samples are sorted by per-sample reconstruction loss and the easiest fraction is selected, and clarify how D_train used in Eq. (14) relates to D_t used for compensation.
  4. [Figure 3(c)] The comparison among 'log2 quantizer', 'DFQ', and 'DFQ(fixed)' lacks axis labels, exact values, and the model/bit-width setting; please add them so the claim that DFQ is comparable to log2-based quantizers can be verified.
  5. [Table 3] The abbreviation 'mACC' is not defined; please state that it is the mean per-class accuracy and define OA as overall accuracy.
  6. [Ablation Studies] The word 'Quantitive' should be 'Quantitative'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: AIQViT fits calibration-set parameters (low-rank weights, ranks, DFQ intervals) and evaluates on held-out test sets; no reported quantity reduces by construction to a fitted input or to a self-citation.

full rationale

The derivation chain is a standard post-training quantization pipeline. The low-rank compensation weights (Eq. 9), the rank search objective (Eqs. 13-14), and the DFQ interval parameters b1 and b2 (Eq. 15) are all optimized using the calibration set and then evaluated on held-out test accuracy (Tables 1-4). No result is defined as the calibration objective itself, and no 'prediction' is a fitted value renamed. The rank search uses a validation split of the calibration set, which is a hyperparameter-selection procedure, not a circular reduction: the reported gains are test-set accuracies after calibration. The only self-citations (Jiang et al. 2022, 2023) appear in the introduction and are not load-bearing for the method. The implausibly low baseline entries (e.g., RepQ-ViT W3/A3 ViT-B at 0.17% in Table 1) raise a serious experimental-validity concern about baseline implementations, but that is a correctness risk rather than a circularity: the paper's own derivation does not assume those baseline values. Therefore no circular step meeting the required evidentiary standard is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on calibration-fit quantities (low-rank matrices, DFQ intervals, architecture parameters) and on several domain assumptions about transfer from calibration to test. The paper provides empirical support but no formal grounding for these assumptions.

free parameters (6)
  • Low-rank compensation matrices A and B per linear layer = not reported (learned on calibration)
    Optimized by Eq. (10) reconstruction loss; the method's gains depend on these fitted weights.
  • DFQ interval bounds b1 and b2 per post-Softmax layer = not reported
    Learned to select the most valuable interval; defines the quantization mapping in Eq. (15).
  • Architecture parameters alpha for rank search = not reported
    Optimized in the bilevel objective of Eqs. (13)-(14); final rank r* = argmax alpha.
  • Curriculum initial sample proportion lambda0 = 0.5
    Set in Optimization Strategy and used in Eq. (17).
  • Candidate rank set S = {10, 20, 50, 100, 150}
    Chosen by hand for all model variants across all tasks.
  • NAS and calibration iteration counts = 2000 NAS, 6000 calibration
    Set in Implementation Details; affects how thoroughly the fitted parameters converge.
assumptions (5)
  • domain assumption The bilevel DARTS approximation (Eqs. 13-14) returns ranks that generalize from the calibration validation split to test data.
    Used to select r for each linear layer; no proof or external validation of the searched architectures.
  • domain assumption Post-Softmax activations' useful information is concentrated in a contiguous interval [b1,b2] per layer, so values outside can be clipped without loss.
    Foundation of DFQ, Eq. (15).
  • domain assumption Block-wise L2 reconstruction loss on calibration data (Eq. 10) is a good proxy for final task accuracy.
    Adopted from BRECQ; no theoretical link established.
  • domain assumption Weight-quantization error in each FC layer is well approximated by a low-rank additive correction.
    Basis of AILoC; inherited from LoRA/QLLM but not proven for quantized ViTs.
  • ad hoc to paper Curriculum learning ordering by reconstruction loss hardness improves optimization for low-bit quantization.
    Introduced in Optimization Strategy; supported only by empirical ablations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers." pith.science (2026). https://pith.science/paper/OT6HH7TF

@misc{pith2026250204628,
  author       = {Pith},
  title        = {Pith review of: AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OT6HH7TF}},
  note         = {Machine review of arXiv:2502.04628}
}
read the original abstract

Post-training quantization (PTQ) has emerged as a promising solution for reducing the storage and computational cost of vision transformers (ViTs). Recent advances primarily target at crafting quantizers to deal with peculiar activations characterized by ViTs. However, most existing methods underestimate the information loss incurred by weight quantization, resulting in significant performance deterioration, particularly in low-bit cases. Furthermore, a common practice in quantizing post-Softmax activations of ViTs is to employ logarithmic transformations, which unfortunately prioritize less informative values around zero. This approach introduces additional redundancies, ultimately leading to suboptimal quantization efficacy. To handle these, this paper proposes an innovative PTQ method tailored for ViTs, termed AIQViT (Architecture-Informed Post-training Quantization for ViTs). First, we design an architecture-informed low rank compensation mechanism, wherein learnable low-rank weights are introduced to compensate for the degradation caused by weight quantization. Second, we design a dynamic focusing quantizer to accommodate the unbalanced distribution of post-Softmax activations, which dynamically selects the most valuable interval for higher quantization resolution. Extensive experiments on five vision tasks, including image classification, object detection, instance segmentation, point cloud classification, and point cloud part segmentation, demonstrate the superiority of AIQViT over state-of-the-art PTQ methods.

Figures

Figures reproduced from arXiv: 2502.04628 by the authors.

Figure 1
Figure 1. The performances of different approaches on dif [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of architecture-informed low-rank com [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) Histogram of the first MHSA module’s post [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of the learned intervals and the in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers

    cs.CV 2025-06 conditional novelty 5.0 of 10

    A two-stage 'activation-first, weights-later' quantization method that reaches competitive 4-bit ViT accuracy with about one epoch of training.

Reference graph

Works this paper leans on

40 extracted references · 34 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Cai, Z.; and Vasconcelos, N. 2018. Cascade R-CNN : Delving Into High Quality Object Detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 6154--6162

    work page 2018

  4. [4]

    Carion, N.; Massa, F.; Synnaeve, G.; Usunier, N.; Kirillov, A.; and Zagoruyko, S. 2020. End-to-End Object Detection with Transformers. In European Conference on Computer Vision, 213--229

    work page 2020

  5. [5]

    Ding, Y.; Qin, H.; Yan, Q.; Chai, Z.; Liu, J.; Wei, X.; and Liu, X. 2022. Towards Accurate Post-Training Quantization for Vision Transformer. In Proceedings of the ACM International Conference on Multimedia, 5380--5388

    work page 2022

  6. [6]

    Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; et al. 2020. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In International Conference on Learning Representations

    2020

  7. [7]

    B.; and Wang, X

    Fang, G.; Ma, X.; Song, M.; Mi, M. B.; and Wang, X. 2023. DepGraph : Towards Any Structural Pruning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 16091--16101

    work page 2023

  8. [8]

    Fang, Y.; Liao, B.; Wang, X.; Fang, J.; Qi, J.; Wu, R.; Niu, J.; and Liu, W. 2021. You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection. In Advances in Neural Information Processing Systems, 26183--26197

    work page 2021

Show all 40 references
  1. [9]

    R.; and Hu, S.-M

    Guo, M.-H.; Cai, J.-X.; Liu, Z.-N.; Mu, T.-J.; Martin, R. R.; and Hu, S.-M. 2021. PCT : Point Cloud Transformer. Computational Visual Media, 7: 187--199

    2021

  2. [10]

    He, K.; Gkioxari, G.; Doll \'a r, P.; and Girshick, R. 2017. Mask R-CNN . In Proceedings of the IEEE International Conference on Computer Vision, 2961--2969

    2017

  3. [11]

    J.; Shen, Y.; Wallis, P.; Allen - Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen - Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W. 2022. LoRA : Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations

    2022

  4. [12]

    Jiang, R.; Yan, Y.; Xue, J.-H.; Chen, S.; Wang, N.; and Wang, H. 2023. Knowledge Distillation Meets Label Noise Learning: Ambiguity-Guided Mutual Label Refinery. IEEE Transactions on Neural Networks and Learning Systems, 1--14

    2023

  5. [13]

    Jiang, R.; Yan, Y.; Xue, J.-H.; Wang, B.; and Wang, H. 2022. When Sparse Neural Network Meets Label Noise Learning: A Multistage Learning Framework. IEEE Transactions on Neural Networks and Learning Systems, 35: 2208--2222

    2022

  6. [14]

    Krizhevsky, A.; Sutskever, I.; and Hinton, G. E. 2012. ImageNet Classification with Deep Convolutional Neural Networks. In Advances in Neural Information Processing Systems

    2012

  7. [15]

    Larsson, G.; Maire, M.; and Shakhnarovich, G. 2016. FractalNet : Ultra-Deep Neural Networks without Residuals. In International Conference on Learning Representations

    2016

  8. [16]

    H.; Miyashita, D.; Chai, E.; Murmann, B.; and Wong, S

    Lee, E. H.; Miyashita, D.; Chai, E.; Murmann, B.; and Wong, S. S. 2017. LogNet : Energy-Efficient Neural Networks Using Logarithmic Computation. In IEEE International Conference on Acoustics, Speech and Signal Processing, 5900--5904

    2017

  9. [17]

    M.; and Shum, H.-Y

    Li, F.; Zhang, H.; Xu, H.; Liu, S.; Zhang, L.; Ni, L. M.; and Shum, H.-Y. 2023 a . Mask DINO : Towards a Unified Transformer-based Framework for Object Detection and Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 3041--3050

    2023

  10. [18]

    Li, Y.; Gong, R.; Tan, X.; Yang, Y.; Hu, P.; Zhang, Q.; Yu, F.; Wang, W.; and Gu, S. 2020. BRECQ: Pushing the Limit of Post-Training Quantization by Block Reconstruction. In International Conference on Learning Representations

    2020

  11. [19]

    Li, Z.; Xiao, J.; Yang, L.; and Gu, Q. 2023 b . Repq-ViT : Scale Reparameterization for Post-Training Quantization of Vision Transformers. In Proceedings of the IEEE International Conference on Computer Vision, 17227--17236

    2023

  12. [20]

    Lin, T.-Y.; Maire, M.; Belongie, S.; Hays, J.; Perona, P.; Ramanan, D.; Doll \'a r, P.; and Zitnick, C. L. 2014. Microsoft COCO : Common Objects in Context. In European Conference on Computer Vision, 740--755

    2014

  13. [21]

    Lin, Y.; Zhang, T.; Sun, P.; Li, Z.; and Zhou, S. 2022. FQ-ViT : Post-Training Quantization for Fully Quantized Vision Transformer. In International Joint Conference on Artificial Intelligence, 1173--1179

    2022

  14. [22]

    Lin, Z.; Courbariaux, M.; Memisevic, R.; and Bengio, Y. 2015. Neural Networks with Few Multiplications. arXiv:1510.03009

    2015 arXiv

  15. [23]

    Liu, H.; Simonyan, K.; and Yang, Y. 2018. DARTS : Differentiable Architecture Search. In International Conference on Learning Representations

    2018

  16. [24]

    Liu, J.; Gong, R.; Wei, X.; Dong, Z.; Cai, J.; and Zhuang, B. 2023. QLLM : Accurate and Efficient Low-Bitwidth Quantization for Large Language Models. In International Conference on Learning Representations

    2023

  17. [25]

    Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; and Guo, B. 2021 a . Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In Proceedings of the IEEE International Conference on Computer Vision, 10012--10022

    2021

  18. [26]

    Liu, Z.; Wang, Y.; Han, K.; Zhang, W.; Ma, S.; and Gao, W. 2021 b . Post-Training Quantization for Vision Transformer. In Advances in Neural Information Processing Systems, 28092--28103

    2021

  19. [27]

    Meng, L.; Li, H.; Chen, B.-C.; Lan, S.; Wu, Z.; Jiang, Y.-G.; and Lim, S.-N. 2022. AdaViT : Adaptive Vision Transformers for Efficient Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 12309--12318

    2022

  20. [28]

    A.; van Baalen, M.; Louizos, C.; and Blankevoort, T

    Nagel, M.; Amjad, R. A.; van Baalen, M.; Louizos, C.; and Blankevoort, T. 2020. Up or Down? Adaptive Rounding for Post-Training Quantization. In International Conference on Machine Learning, 7197--7206

    2020

  21. [29]

    Nagel, M.; Fournarakis, M.; Bondarenko, Y.; and Blankevoort, T. 2022. Overcoming Oscillations in Quantization-Aware Training. In International Conference on Machine Learning, 16318--16330

    2022

  22. [30]

    Touvron, H.; Cord, M.; Douze, M.; Massa, F.; Sablayrolles, A.; and J \'e gou, H. 2021. Training Data-Efficient Image Transformers & Distillation Through Attention. In International Conference on Machine Learning, 10347--10357

    2021

  23. [31]

    Wang, X.; Chen, Y.; and Zhu, W. 2022. A Survey on Curriculum Learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9): 4555--4576

    2022

  24. [32]

    Wei, X.; Gong, R.; Li, Y.; Liu, X.; and Yu, F. 2021. QDrop: Randomly Dropping Quantization for Extremely Low-bit Post-Training Quantization. In International Conference on Learning Representations

    2021

  25. [33]

    Wu, Z.; Song, S.; Khosla, A.; Yu, F.; Zhang, L.; Tang, X.; and Xiao, J. 2015. 3D ShapeNets : A Deep Representation for Volumetric Shapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 1912--1920

    2015

  26. [34]

    Xiao, Y.; Liu, A.; Zhang, T.; Qin, H.; Guo, J.; and Liu, X. 2023. RobustMQ: Benchmarking Robustness of Quantized Models. Visual Intelligence, 1(1): 30

    2023

  27. [35]

    Yan, P.; Liu, X.; Zhang, P.; and Lu, H. 2023. Learning Convolutional Multi-level Transformers for Image-based Person Re-identification. Visual Intelligence, 1(1): 24

    2023

  28. [36]

    G.; Ceylan, D.; Shen, I.-C.; Yan, M.; Su, H.; Lu, C.; Huang, Q.; Sheffer, A.; and Guibas, L

    Yi, L.; Kim, V. G.; Ceylan, D.; Shen, I.-C.; Yan, M.; Su, H.; Lu, C.; Huang, Q.; Sheffer, A.; and Guibas, L. 2016. A Scalable Active Framework for Region Annotation in 3D Shape Collections. ACM Transactions on Graphics, 35(6): 1--12

    2016

  29. [37]

    Yuan, Z.; Xue, C.; Chen, Y.; Wu, Q.; and Sun, G. 2022. PTQ4ViT : Post-Training Quantization for Vision Transformers with Twin Uniform Quantization. In European Conference on Computer Vision, 191--207

    2022

  30. [38]

    Zhao, B.; Cui, Q.; Song, R.; Qiu, Y.; and Liang, J. 2022. Decoupled Knowledge Distillation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 11953--11962

    2022

  31. [39]

    H.; and Koltun, V

    Zhao, H.; Jiang, L.; Jia, J.; Torr, P. H.; and Koltun, V. 2021. Point Transformer. In Proceedings of the IEEE International Conference on Computer Vision, 16259--16268

    2021

  32. [40]

    Zhong, Y.; Hu, J.; Lin, M.; Chen, M.; and Ji, R. 2023. I&S-ViT : An Inclusive & Stable Method for Pushing the Limit of Post-Training ViTs Quantization. arXiv:2311.10126

    2023

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.