Pith. sign in

REVIEW 3 major objections 2 minor 55 references

FACT: A Simple and Efficient Framework for Active Finetuning

T0 review · 3 major / 2 minor · reviewed 2026-06-28 · grok-4.3

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

Pith's one-line read A three-phase hierarchical finetuning framework improves active finetuning by avoiding feature distortion in pretrained models with scarce labels.

desk verdict FACT introduces a three-phase finetuning approach for active learning in vision to limit pretrained feature distortion and reports large gains on ViT, but provides no direct test that the phases are what drive the improvement. read the letter →

arxiv 2606.02079 v1 pith:DST6VGS7 submitted 2026-06-01 cs.CV

Wenshuai Xu , You Song , Yuzhuo Cui , Minjie Ren , Qingjie Liu , Zhenghui Hu This is my paper
classification cs.CV
keywords activefinetuninglearningframeworkvisiontransformerimageclassificationfeaturedistortiondataefficiencyfrozenaugmentation
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

The paper aims to show that full finetuning distorts pretrained features during active learning when the model is large relative to the amount of selected data. It proposes FACT as a three-phase hierarchical strategy that adapts the model more carefully while using frozen feature augmentation. This approach would matter if it enables better performance from large vision models without the overfitting that typically occurs in low-data active selection settings. Experiments test the method across multiple datasets, architectures, and sampling ratios to demonstrate gains exceeding 20 percent on ViT models for CIFAR and ImageNet benchmarks.

What carries the argument

The three-phase hierarchical finetuning strategy with frozen feature augmentation (FroFA), which progressively adapts the model to actively selected data without full parameter updates that would distort pretrained features.

What would settle it

An experiment in which standard full finetuning on the same actively selected low-ratio subsets for ViT on ImageNet-1k matches or exceeds FACT's reported accuracy.

Watch

Extended reading notes

Core claim

The paper claims that full finetuning inevitably distorts pretrained features due to distribution shift in active finetuning, and that a three-phase hierarchical finetuning framework called FACT, which incorporates frozen feature augmentation strategies, can systematically mitigate this distortion while maintaining parameter efficiency and achieving over 20 percent performance gains on ViT models for CIFAR10, CIFAR100, and ImageNet-1k under low sampling ratios.

Load-bearing premise

The three-phase hierarchical strategy can avoid feature distortion from full finetuning without creating its own overfitting or selection biases.

Editorial extensions

If this is right

  • Higher accuracy on image classification tasks when only a small fraction of labels is available.
  • Consistent gains across convolutional, transformer, and LSTM-based vision architectures.
  • Effective results on both balanced, imbalanced, and fine-grained datasets.
  • Better parameter efficiency than full finetuning while reaching new state-of-the-art active finetuning performance.

Reading between the lines

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

  • The same phased approach might reduce feature distortion in other low-data transfer learning scenarios outside active selection.
  • Interactions between the specific data selection method and the finetuning phases could be studied to further improve results.
  • Testing the framework on tasks beyond image classification could indicate whether the benefit generalizes to other modalities.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 2 minor

Summary. The paper introduces the FiAF task (active finetuning with emphasis on finetuning methodology) and proposes FACT, a three-phase hierarchical finetuning framework that uses frozen feature augmentation (FroFA) to mitigate pretrained feature distortion under low-data active learning regimes. Experiments cover CIFAR-10/100, ImageNet-1k, long-tailed and fine-grained datasets, multiple architectures (ConvNeXt, ViT, ViL), and 3-5 sampling ratios; the central empirical claim is >20% gains on ViT for CIFAR-10/100 and ImageNet-1k at low sampling ratios while preserving parameter efficiency.

Significance. If the performance deltas are reproducible and the mechanism is isolated, the work would supply a simple, parameter-efficient alternative to full finetuning in active-learning pipelines for large vision models, directly addressing the feature-distortion problem that becomes acute when labeled data is scarce relative to model capacity.

major comments (3)
  1. [Abstract] Abstract and experimental sections: the >20% gains on ViT for CIFAR-10/100/ImageNet-1k are stated without error bars, exact dataset splits, number of runs, or ablation tables that isolate the three-phase hierarchy from the active selector and FroFA augmentations; without these the contribution of distortion avoidance cannot be evaluated.
  2. [Experiments] No section supplies a direct metric (feature-space distance, activation drift, or linear-probe accuracy on held-out data) comparing distortion levels of FACT versus full finetuning on identical selected subsets; the performance delta could therefore be driven by the selector, augmentations, or phase-specific regularization rather than the claimed mechanism.
  3. [Method] The premise that the three-phase hierarchy systematically avoids feature distortion is load-bearing for the central claim, yet the manuscript provides no controlled comparison that holds the selected subset fixed while varying only the finetuning strategy.
minor comments (2)
  1. Notation for the three phases and FroFA variants is introduced without a compact summary table or diagram that would allow readers to track which components are frozen versus updated in each phase.
  2. Dataset splits, sampling ratios, and hyper-parameter choices for the active selector are described at high level; explicit values or references to standard splits would improve reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on improving experimental rigor and isolating the mechanism. We address each major comment below and will make targeted revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract and experimental sections: the >20% gains on ViT for CIFAR-10/100/ImageNet-1k are stated without error bars, exact dataset splits, number of runs, or ablation tables that isolate the three-phase hierarchy from the active selector and FroFA augmentations; without these the contribution of distortion avoidance cannot be evaluated.

    Authors: We agree that reproducibility details are important. The experiments were conducted with 3 random seeds; we will add error bars, specify the exact train/validation splits, and expand the ablation tables in Section 4 to more clearly separate the contributions of the three-phase hierarchy, active selector, and FroFA. These changes will be included in the revision. revision: yes

  2. Referee: [Experiments] No section supplies a direct metric (feature-space distance, activation drift, or linear-probe accuracy on held-out data) comparing distortion levels of FACT versus full finetuning on identical selected subsets; the performance delta could therefore be driven by the selector, augmentations, or phase-specific regularization rather than the claimed mechanism.

    Authors: This observation is correct; the current manuscript relies on end-task performance rather than an explicit distortion metric. To address it, we will add linear-probe accuracy results on held-out data comparing FACT and full finetuning using the same selected subsets, providing direct evidence for the distortion-avoidance claim. revision: yes

  3. Referee: [Method] The premise that the three-phase hierarchy systematically avoids feature distortion is load-bearing for the central claim, yet the manuscript provides no controlled comparison that holds the selected subset fixed while varying only the finetuning strategy.

    Authors: Our primary results already compare FACT against full finetuning while keeping the active selector and sampling ratios identical across methods, which holds data selection fixed. To make the isolation more explicit, we will add a dedicated controlled ablation subsection that further varies only the finetuning strategy on fixed subsets. revision: partial

Circularity Check

0 steps flagged · score 0.0 of 10

Empirical framework proposal with no derivation chain or self-referential reductions

full rationale

The paper introduces the FiAF task and proposes the FACT three-phase hierarchical finetuning framework, supported by experiments on multiple datasets, architectures, and sampling ratios. No equations, parameter fits presented as predictions, uniqueness theorems, or self-citations appear in the provided text. Performance gains are reported from direct empirical evaluation rather than any derivation that reduces to its own inputs by construction. The central claims rest on experimental results, not on a mathematical chain that could exhibit circularity. This is a standard empirical contribution in active learning without load-bearing self-referential steps.

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

No free parameters, axioms, or invented entities are described in the abstract; the ledger is empty because the text supplies no technical details on modeling choices or assumptions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FACT: A Simple and Efficient Framework for Active Finetuning." pith.science (2026). https://pith.science/paper/DST6VGS7

@misc{pith2026260602079,
  author       = {Pith},
  title        = {Pith review of: FACT: A Simple and Efficient Framework for Active Finetuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DST6VGS7}},
  note         = {Machine review of arXiv:2606.02079}
}
read the original abstract

The main goal of active finetuning is to improve a pretrained model's performance on a specific task or domain by finetuning it with carefully selected informative or challenging data. Previous research has predominantly focused on the active aspect (i.e., data selection) while uniformly employing full finetuning for model adaptation, which inevitably distorts pretrained features due to distribution shift. This issue becomes particularly pronounced when the model size is large relative to the finetuning data quantity, leading to heightened overfitting risks. To address this critical gap, we formally outline the FiAF task that emphasizes systematic exploration of finetuning methodologies in active learning. We propose FACT, a three-phase hierarchical finetuning framework featuring both efficiency and simplicity, specifically designed for active finetuning scenarios. Our comprehensive experiments span: (1) Three major dataset categories encompassing classic (CIFAR10, CIFAR100, ImageNet-1k), imbalanced (CIFAR10-LT, CIFAR100-LT), and fine-grained (StanfordCars, FGVCAircraft) image classification datasets, each evaluated under 3-5 distinct sampling ratios; (2) Diverse pretrained architectures including Convolutional Neural Network (ConvNeXt), Vision Transformer (ViT), and Vision LSTM (ViL) networks; (3) A systematic investigation of frozen feature augmentation (FroFA) strategies. (4) A comprehensive and rigorous analysis of efficiency and generalizability. The results demonstrate significant improvements with strong generalization and robustness. Notably, under low sampling ratios, our framework achieves remarkable performance gains of over 20% on the ViT model for CIFAR10, CIFAR100, and ImageNet-1k benchmarks. This systematic approach establishes new state-of-the-art performance while maintaining parameter efficiency, proving particularly effective when labeled data is scarce.

Figures

Figures reproduced from arXiv: 2606.02079 by the authors.

Figure 1
Figure 1. We focus on the Finetuning task in Active Finetuning (FiAF). Prior studies have emphasized the active (i.e., data selection) over the finetuning. However, the full finetuning in FiAF faces overfitting challenges due to small￾scale data and large pretrained vision model (PVM). The compatibility of finetuning methods in the active finetuning scheme has often been overlooked in previous research. Although the pretraini… view at source ↗
Figure 2
Figure 2. The proposed FACT framework is specifically designed for the FiAF task to address the problems of data distribution shift and over-parameterization of pretrained models during the finetuning phase. We highlight it with a light blue background in the figure, involves a three-phase process: (1) Linear Probing (LP): In this phase, the parameters of the pretrained model are frozen. Only the linear classifier, initialize… view at source ↗
Figure 3
Figure 3. L3FACT : LP-LoRA-LM Finetuning in ACTive finetuning. From the perspective of optimizing parameter efficiency, we propose a variant of the FACT method, L3FACT, by replacing the second phase of full finetuning in the finetuning phase shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Classification accuracy of methods applied on different backbones on CIFAR10, CIFAR100, and ImageNet-1k. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 5 canonical work pages

  1. [1]

    Bridging the gap between pre-training and fine-tuning for end-to-end speech translation,

    C. Wang, Y . Wu, S. Liu, Z. Yang, and M. Zhou, “Bridging the gap between pre-training and fine-tuning for end-to-end speech translation,” inProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020, pp. 9161–9168

    2020

  2. [2]

    Code: contrastive pre-training with adversarial fine-tuning for zero-shot expert linking,

    B. Chen, J. Zhang, X. Zhang, X. Tang, H. Chen, C. Li, P. Zhang, J. Tang et al., “Code: contrastive pre-training with adversarial fine-tuning for zero-shot expert linking,” inProceedings of the AAAI Conference on Artificial Intelligence (AAAI), vol. 36, 2022, pp. 11 846–11 854

    2022

  3. [3]

    Equi-tuning: Group equivariant fine-tuning of pretrained models,

    S. Basu, P. Sattigeri, K. N. Ramamurthy, V . Chenthamarakshan, K. R. Varshney, L. R. Varshney, and P. Das, “Equi-tuning: Group equivariant fine-tuning of pretrained models,” inProceedings of the AAAI Confer- ence on Artificial Intelligence (AAAI), vol. 37, 2023, pp. 6788–6796

    2023

  4. [4]

    Understanding Uncertainty Sampling via Equivalent Loss

    S. Liu and X. Li, “Understanding uncertainty sampling,”arXiv preprint arXiv:2307.02719, 2023

    work page Pith review arXiv 2023

  5. [5]

    Learning loss for active learning,

    D. Yoo and I. S. Kweon, “Learning loss for active learning,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 93–102

    2019

  6. [6]

    Deep batch active learning by diverse, uncertain gradient lower bounds,

    J. T. Ash, C. Zhang, A. Krishnamurthy, J. Langford, and A. Agar- wal, “Deep batch active learning by diverse, uncertain gradient lower bounds,” inInternational Conference on Learning Representations (ICLR), 2020

    2020

  7. [7]

    Task-aware variational adversarial active learning,

    K. Kim, D. Park, K. I. Kim, and S. Y . Chun, “Task-aware variational adversarial active learning,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 8166–8175

    2021

  8. [8]

    Active learning by feature mixing,

    A. Parvaneh, E. Abbasnejad, D. Teney, G. R. Haffari, A. Van Den Hen- gel, and J. Q. Shi, “Active learning by feature mixing,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 12 237–12 246

    2022

Show all 55 references
  1. [9]

    Deep active learning with noise stability,

    X. Li, P. Yang, Y . Gu, X. Zhan, T. Wang, M. Xu, and C. Xu, “Deep active learning with noise stability,” inProceedings of the AAAI Conference on Artificial Intelligence (AAAI), vol. 38, 2024, pp. 13 655–13 663

    2024

  2. [10]

    Re- ducing label effort: Self-supervised meets active learning,

    J. Z. Bengar, J. van de Weijer, B. Twardowski, and B. Raducanu, “Re- ducing label effort: Self-supervised meets active learning,” inIEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 1631– 1639

    2021

  3. [11]

    Active finetuning: Exploiting annotation budget in the pretraining-finetuning paradigm,

    Y . Xie, H. Lu, J. Yan, X. Yang, M. Tomizuka, and W. Zhan, “Active finetuning: Exploiting annotation budget in the pretraining-finetuning paradigm,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 23 715–23 724

    2023

  4. [12]

    Activedc: Distribution calibration for active finetuning,

    W. Xu, Z. Hu, Y . Lu, J. Meng, Q. Liu, and Y . Wang, “Activedc: Distribution calibration for active finetuning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 16 996–17 005

    2024

  5. [13]

    Boundary matters: A bi-level active finetuning method,

    H. Lu, Y . Xie, X. Yang, and J. Yan, “Boundary matters: A bi-level active finetuning method,”Advances in Neural Information Processing Systems, vol. 37, pp. 35 945–35 972, 2024

    2024

  6. [14]

    Vecaf: Vision-language collaborative active finetuning with training objective awareness,

    R. Zhang, Z. Cai, H. Yang, Z. Liu, D. Gudovskiy, T. Okuno, Y . Nakata, K. Keutzer, B. Chang, Y . Duet al., “Vecaf: Vision-language collaborative active finetuning with training objective awareness,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp...

    2024

  7. [15]

    Fine-tuning can distort pretrained features and underperform out-of- distribution,

    A. Kumar, A. Raghunathan, R. M. Jones, T. Ma, and P. Liang, “Fine-tuning can distort pretrained features and underperform out-of- distribution,” inInternational Conference on Learning Representations (ICLR), 2022

    2022

  8. [16]

    Revisit finetuning strategy for few-shot learning to transfer the emdeddings,

    H. Wang, T. Yue, X. Ye, Z. He, B. Li, and Y . Li, “Revisit finetuning strategy for few-shot learning to transfer the emdeddings,” inInterna- tional Conference on Learning Representations (ICLR), 2023

    2023

  9. [17]

    Which pretrain sam- ples to rehearse when finetuning pretrained models?

    A. Bai, C.-K. Yeh, C.-J. Hsieh, and A. Taly, “Which pretrain sam- ples to rehearse when finetuning pretrained models?”arXiv preprint arXiv:2402.08096, 2024

    2024

  10. [18]

    Deep neural networks for high dimension, low sample size data

    B. Liu, Y . Wei, Y . Zhang, and Q. Yang, “Deep neural networks for high dimension, low sample size data.” inInternational Joint Conference on Artificial Intelligence (IJCAI), vol. 2017, 2017, pp. 2287–2293

    2017

  11. [19]

    Pushing the limits of simple pipelines for few-shot learning: External data and fine- tuning make a difference,

    S. X. Hu, D. Li, J. St ¨uhmer, M. Kim, and T. M. Hospedales, “Pushing the limits of simple pipelines for few-shot learning: External data and fine- tuning make a difference,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. ...

    2022

  12. [20]

    Active learning on a budget: Opposite strategies suit high and low budgets,

    G. Hacohen, A. Dekel, and D. Weinshall, “Active learning on a budget: Opposite strategies suit high and low budgets,” inInternational Confer- ence on Machine Learning (ICML), 2022, pp. 8175–8195

    2022

  13. [21]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational Conference on Machine Learning (ICML), 2021, pp. 8748–8763

    2021

  14. [22]

    Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,” inInternational Conference on Machine Learning (ICML). PMLR, 2023, pp. 19 730–19 742

    2023

  15. [23]

    An image is worth 16x16 words: Trans- formers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Trans- formers for image recognition at scale,” inInternational Conference on Learnin...

    2021

  16. [24]

    Training data-efficient image transformers & distillation through attention,

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. J ´egou, “Training data-efficient image transformers & distillation through attention,” inInternational Conference on Machine Learning (ICML), 2021, pp. 10 347–10 357

    2021

  17. [25]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  18. [26]

    Vision-LSTM: xLSTM as generic vision backbone,

    B. Alkin, M. Beck, K. P ¨oppel, S. Hochreiter, and J. Brandstetter, “Vision-LSTM: xLSTM as generic vision backbone,” inThe Thirteenth International Conference on Learning Representations (ICLR), 2025

    2025

  19. [27]

    A closer look at few-shot classification,

    W.-Y . Chen, Y .-C. Liu, Z. Kira, Y .-C. Wang, and J.-B. Huang, “A closer look at few-shot classification,” inInternational Conference on Learning Representations (ICLR), 2019

    2019

  20. [28]

    Charting the right manifold: Manifold mixup for few-shot learning,

    P. Mangla, N. Kumari, A. Sinha, M. Singh, B. Krishnamurthy, and V . N. Balasubramanian, “Charting the right manifold: Manifold mixup for few-shot learning,” inProceedings of the IEEE/CVF winter conference on applications of computer vision (WACV), 2020, pp. 2218–2227

    2020

  21. [29]

    Autoaug- ment: Learning augmentation strategies from data,

    E. D. Cubuk, B. Zoph, D. Mane, V . Vasudevan, and Q. V . Le, “Autoaug- ment: Learning augmentation strategies from data,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2019

    2019

  22. [30]

    Trivialaugment: Tuning-free yet state-of- the-art data augmentation,

    S. G. Muller and F. Hutter, “Trivialaugment: Tuning-free yet state-of- the-art data augmentation,” in2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021

    2021

  23. [31]

    Partial is better than all: Revisiting fine-tuning strategy for few-shot learning,

    Z. Shen, Z. Liu, J. Qin, M. Savvides, and K.-T. Cheng, “Partial is better than all: Revisiting fine-tuning strategy for few-shot learning,” inProceedings of the AAAI conference on artificial intelligence (AAAI), vol. 35, 2021, pp. 9594–9602

    2021

  24. [32]

    LoRA+: Efficient low rank adaptation of large models,

    S. Hayou, N. Ghosh, and B. Yu, “LoRA+: Efficient low rank adaptation of large models,” inInternational Conference on Machine Learning (ICML), 2024

    2024

  25. [33]

    Lora-ga: Low-rank adaptation with gradient approximation,

    S. Wang, L. Yu, and J. Li, “Lora-ga: Low-rank adaptation with gradient approximation,”arXiv preprint arXiv:2407.05000, 2024

    2024

  26. [34]

    Qlora: Efficient finetuning of quantized llms,

    T. Dettmers, A. Pagnoni, A. Holtzman, and L. Zettlemoyer, “Qlora: Efficient finetuning of quantized llms,”Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  27. [35]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, yelong shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inInternational Conference on Learning Representations (ICLR), 2022

    2022

  28. [36]

    Parameter-efficient fine-tuning for large models: A comprehensive survey,

    Z. Han, C. Gao, J. Liu, J. Zhang, and S. Q. Zhang, “Parameter-efficient fine-tuning for large models: A comprehensive survey,” 2024. [Online]. Available: https://arxiv.org/abs/2403.14608

    2024 arXiv

  29. [37]

    Visual prompt tuning,

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim, “Visual prompt tuning,” inEuropean conference on computer vision (ECCV). Springer, 2022, pp. 709–727

    2022

  30. [38]

    Prefix-tuning: Optimizing continuous prompts for generation,

    X. L. Li and P. Liang, “Prefix-tuning: Optimizing continuous prompts for generation,” inProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2...

    2021

  31. [39]

    Parameter- efficient fine-tuning in spectral domain for point cloud learning,

    D. Liang, T. Feng, X. Zhou, Y . Zhang, Z. Zou, and X. Bai, “Parameter- efficient fine-tuning in spectral domain for point cloud learning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  32. [40]

    Finetune like you pretrain: Improved finetuning of zero-shot vision models,

    S. Goyal, A. Kumar, S. Garg, Z. Kolter, and A. Raghunathan, “Finetune like you pretrain: Improved finetuning of zero-shot vision models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 19 338–19 347

    2023

  33. [41]

    Frozen feature augmentation for few-shot image classification,

    A. B ¨ar, N. Houlsby, M. Dehghani, and M. Kumar, “Frozen feature augmentation for few-shot image classification,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 16 046–16 057

    2024

  34. [42]

    Big transfer (bit): General visual representation learning,

    A. Kolesnikov, L. Beyer, X. Zhai, J. Puigcerver, J. Yung, S. Gelly, and N. Houlsby, “Big transfer (bit): General visual representation learning,” inEuropean Conference on Computer Vision (ECCV), 2020, pp. 491– 507. IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 00, NO. 0, MAY 2025 12

    2020

  35. [43]

    Scaling vision transformers,

    X. Zhai, A. Kolesnikov, N. Houlsby, and L. Beyer, “Scaling vision transformers,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2022, pp. 12 104–12 113

    2022

  36. [44]

    A progressive batching l-bfgs method for machine learning,

    R. Bollapragada, J. Nocedal, D. Mudigere, H.-J. Shi, and P. T. P. Tang, “A progressive batching l-bfgs method for machine learning,” in International Conference on Machine Learning (ICML). PMLR, 2018, pp. 620–629

    2018

  37. [45]

    An accelerated linearly convergent stochastic l-bfgs algorithm,

    D. Chang, S. Sun, and C. Zhang, “An accelerated linearly convergent stochastic l-bfgs algorithm,”IEEE Transactions on Neural Networks and Learning Systems, vol. 30, no. 11, pp. 3338–3346, 2019

    2019

  38. [46]

    Set trans- former: A framework for attention-based permutation-invariant neural networks,

    J. Lee, Y . Lee, J. Kim, A. Kosiorek, S. Choi, and Y . Teh, “Set trans- former: A framework for attention-based permutation-invariant neural networks,”International Conference on Machine Learning (ICML), 2019

    2019

  39. [47]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hintonet al., “Learning multiple layers of features from tiny images,” 2009

    2009

  40. [48]

    Imagenet large scale visual recognition challenge,

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernsteinet al., “Imagenet large scale visual recognition challenge,”International Journal of Computer Vision (IJCV), vol. 115, pp. 211–252, 2015

    2015

  41. [49]

    Learning imbal- anced datasets with label-distribution-aware margin loss,

    K. Cao, C. Wei, A. Gaidon, N. Arechiga, and T. Ma, “Learning imbal- anced datasets with label-distribution-aware margin loss,” inAdvances in Neural Information Processing Systems (NeurIPS), 2019

    2019

  42. [50]

    3d object representations for fine-grained categorization,

    J. Krause, M. Stark, J. Deng, and L. Fei-Fei, “3d object representations for fine-grained categorization,” inInternational IEEE Workshop on 3D Representation and Recognition, 2013

    2013

  43. [51]

    Fine- grained visual classification of aircraft,

    S. Maji, E. Rahtu, J. Kannala, M. Blaschko, and A. Vedaldi, “Fine- grained visual classification of aircraft,”arXiv preprint arXiv:1306.5151, 2013

    2013 arXiv

  44. [52]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” inIEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 9650–9660

    2021

  45. [53]

    MMSegmentation: Openmmlab semantic seg- mentation toolbox and benchmark,

    M. Contributors, “MMSegmentation: Openmmlab semantic seg- mentation toolbox and benchmark,” https://github.com/open-mmlab/ mmsegmentation, 2020

    2020

  46. [54]

    Unified perceptual parsing for scene understanding,

    T. Xiao, Y . Liu, B. Zhou, Y . Jiang, and J. Sun, “Unified perceptual parsing for scene understanding,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 418–434

    2018

  47. [55]

    Scene parsing through ade20k dataset,

    B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba, “Scene parsing through ade20k dataset,” inProceedings of the IEEE conference on computer vision and pattern recognition (CVPR), 2017, pp. 633–641. Wenshuai Xureceived the B.S. degree in soft- ware engineering ...

    2017

Pith tools

Reviewed June 28, 2026 · model on record in the stance chip above.