pith. sign in

arxiv: 2605.23198 · v1 · pith:IKQXU7ZInew · submitted 2026-05-22 · 💻 cs.LG

Label-Efficient Dataset Pruning via Semi-Supervised Pseudo-Labeling

Yeseul Cho , Baekrok Shin , Changmin Kang , Chulhee Yun This is my paper

Pith reviewed 2026-05-25 05:23 UTC · model grok-4.3

classification 💻 cs.LG
keywords dataset pruningsemi-supervised learningpseudo-labelingcoreset selectionlabel-efficient learningexample difficultytraining dynamics
0
0 comments X

The pith

A small randomly labeled subset and semi-supervised pseudo-labeling lets supervised pruning methods select reliable coresets from mostly unlabeled data.

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

The paper proposes SemiPrune to perform dataset pruning when only a small fraction of data has labels. It runs semi-supervised learning on the labeled subset to assign pseudo-labels to the rest, then applies existing supervised pruning techniques that rely on training dynamics to estimate example difficulty and pick an informative coreset. This approach learns directly from the target data rather than from features of a pretrained model, which can mismatch the target distribution. The result is a label-efficient method that works on domain-specific, corrupted, and long-tailed image datasets while matching or exceeding prior label-free baselines.

Core claim

SemiPrune generates pseudo-labels for unlabeled examples by training a semi-supervised model on a small randomly chosen labeled subset, then uses the resulting pseudo-labeled pool to compute training dynamics that indicate example difficulty, and finally selects a coreset with any supervised pruning method.

What carries the argument

Pseudo-label-induced training dynamics for difficulty estimation and coreset selection after semi-supervised learning on a small labeled subset.

If this is right

  • Existing supervised pruning algorithms can now be applied directly to largely unlabeled pools without modification.
  • Pruning performance improves on datasets whose distribution differs from common pretraining data.
  • Annotation budgets can be reduced while still producing competitive coresets on standard, corrupted, and long-tailed image benchmarks.
  • Difficulty signals come from dynamics on the target distribution rather than external features.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The initial small labeled subset could itself be chosen actively rather than randomly to further improve pseudo-label quality.
  • Iterative pseudo-label refinement during the pruning process might reduce error propagation from early mistakes.
  • The same pseudo-label pipeline could be tested on non-image modalities where pretrained models are even less reliable.

Load-bearing premise

The pseudo-labels produced by semi-supervised learning on the small labeled subset are accurate enough that the resulting training dynamics reliably indicate example difficulty for coreset selection.

What would settle it

On a domain-shifted dataset, train models on coresets chosen by SemiPrune versus by a pretrained-feature baseline and check whether the SemiPrune coreset yields lower test accuracy.

Figures

Figures reproduced from arXiv: 2605.23198 by Baekrok Shin, Changmin Kang, Chulhee Yun, Yeseul Cho.

Figure 1
Figure 1. Figure 1: t-SNE visualization of Food-101 embeddings from ImageNet-pretrained DINO (a) and SSL-trained models (b). For each feature space, panels are colored by ground-truth labels. depends on whether the pretrained feature space is well aligned with the semantic structure of the target dataset. First, when the target dataset is weakly aligned with the pretraining data, deep clustering may fail to recover semantical… view at source ↗
Figure 2
Figure 2. Figure 2: Effect of the initial label budget on coreset performance. We compare our method with Score Extrapolation under identical initial label budgets and include label-free baselines for reference. We next study the effect of the initial label budget on coreset performance. Since both our method and Score Extrapolation rely on a small annotated subset, we compare them under the same label budgets. As shown in [… view at source ↗
Figure 3
Figure 3. Figure 3: t-SNE visualization of corrupted CIFAR-100 embeddings from ImageNet-pretrained DINO (a) and SSL-trained models (b). For each feature space, panels are colored by ground-truth labels. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: t-SNE visualization of long-tailed CIFAR-100 with an imbalance factor of 0.1 embeddings from ImageNet-pretrained DINO (a) and SSL-trained models (b). For each feature space, panels are colored by ground-truth labels. (a) Embeddings from DINO (b) Embeddings from SSL-trained model [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: t-SNE visualization of Caltech-101 embeddings from ImageNet-pretrained DINO (a) and SSL-trained models (b). For each feature space, panels are colored by ground-truth labels [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Class-wise distribution of assigned pseudo-labels. Deep-Clustering does not distribute pseudo-label assignments evenly across classes, leaving some classes with no assigned samples at all, whereas Semi-Supervised Learning yields a relatively more balanced distribution. For corrupted samples, both DC and SSL assign a substantial portion of samples to the same class. However, SSL maintains a relatively broad… view at source ↗
Figure 7
Figure 7. Figure 7: Class-wise distribution of assigned pseudo-labels. 0 20 40 60 80 100 Class Index 0 100 200 300 400 500 Number of Samples DC Pseudo-labels Ground Truth DC 0 20 40 60 80 100 Class Index SSL Pseudo-labels Ground Truth SSL Distribution of Pseudo Labels on CIFAR-100-LT-IF0.1 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Class-wise distribution of assigned pseudo-labels. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Class-wise distribution of assigned pseudo-labels [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Effect of the initial label budget on coreset performance on CIFAR-100. 1% 2% 5% 10% 73.2 74.5 75.8 77.1 78.4 Accuracy 30% Pruning 1% 2% 5% 10% 61.9 65.3 68.7 72.2 75.6 50% Pruning 1% 2% 5% 10% 57.7 60.8 63.9 67.0 70.1 70% Pruning 1% 2% 5% 10% 52.0 55.4 58.8 62.3 65.7 80% Pruning 1% 2% 5% 10% 34.8 40.1 45.4 50.7 56.0 90% Pruning FOOD-101 Initial Label Budget Fully Supervised Random ELFS (DINO) ELFS (Self)… view at source ↗
Figure 11
Figure 11. Figure 11: Effect of the initial label budget on coreset performance on Food-101. C.2 Coreset Performance Comparison Under Same Annotation Budget We compare coreset performance under a fixed annotation budget on Caltech-101. Semi-AUM+Cutoff first includes the randomly labeled 10% subset used for semi-supervised learning. It then ranks the remaining pseudo-labeled examples by their Semi-AUM scores and selects additio… view at source ↗
read the original abstract

Dataset pruning reduces the storage and training costs of deep learning by selecting an informative subset from a large dataset. However, most existing pruning methods require fully labeled data, which limits their applicability in realistic settings where unlabeled data are abundant and annotation is costly. Recent label-free pruning methods address this issue, but they rely on features from pretrained models to estimate example difficulty. This dependence can be unreliable when the target dataset differs substantially from the pretraining distribution. We propose SemiPrune, a label-efficient dataset pruning framework, using only a small randomly labeled subset, that uses semi-supervised learning to generate pseudo-labels for unlabeled data, allowing existing supervised pruning methods that require label information to be seamlessly applied to the resulting pseudo-labeled training pool. We then estimate example difficulty from pseudo-label-induced training dynamics and select a coreset. By learning directly from the target dataset, our method better captures the target distribution and provides more reliable signals for difficulty estimation and coreset selection. We validate our approach on domain-specific, image-corrupted, and long-tailed datasets, where it achieves state-of-the-art performance among label-free and label-efficient baselines, while also demonstrating competitive performance on standard benchmarks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes SemiPrune, a label-efficient dataset pruning method that uses a small randomly labeled subset to train a semi-supervised model, generates pseudo-labels for the remaining data, and then applies existing supervised pruning techniques (based on training dynamics) to select a coreset. It claims this captures the target distribution better than pretrained-feature methods and achieves SOTA among label-free/label-efficient baselines on domain-specific, corrupted, and long-tailed datasets while remaining competitive on standard benchmarks.

Significance. If the central mechanism holds, the work would meaningfully extend dataset pruning to realistic low-label regimes without relying on potentially mismatched pretrained models, directly addressing annotation cost and distribution shift issues. The approach re-uses existing supervised difficulty estimators on a pseudo-labeled pool, which is a pragmatic strength if the induced dynamics remain faithful.

major comments (2)
  1. [§3 (method description)] The central claim (abstract and §3) that 'pseudo-label-induced training dynamics' reliably indicate example difficulty for coreset selection rests on an unexamined assumption: that SSL pseudo-labels from a small random labeled subset produce monotonic loss/gradient signals with respect to true difficulty. No derivation, bound, or even correlation analysis is provided showing this holds under domain shift, corruption, or long-tailed distributions, where confirmation bias in SSL is known to distort per-example losses.
  2. [§4 (experimental validation)] Experiments (presumably §4) report SOTA performance on the highlighted regimes but provide no validation of pseudo-label accuracy, no ablations on labeled-subset size, and no comparison of difficulty rankings before/after pseudo-labeling. Without these, it is impossible to attribute gains to the proposed mechanism rather than incidental factors.
minor comments (1)
  1. [§3] Notation for the semi-supervised component and the difficulty estimator should be introduced with explicit equations rather than prose descriptions to allow reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments point-by-point below and will revise the manuscript accordingly to strengthen the presentation of the method and experiments.

read point-by-point responses

  1. Referee: [§3 (method description)] The central claim (abstract and §3) that 'pseudo-label-induced training dynamics' reliably indicate example difficulty for coreset selection rests on an unexamined assumption: that SSL pseudo-labels from a small random labeled subset produce monotonic loss/gradient signals with respect to true difficulty. No derivation, bound, or even correlation analysis is provided showing this holds under domain shift, corruption, or long-tailed distributions, where confirmation bias in SSL is known to distort per-example losses.

    Authors: We agree that the current manuscript provides no theoretical derivation or bound and does not include an explicit correlation analysis between pseudo-label-induced difficulty scores and ground-truth difficulty scores. The empirical results on domain-specific, corrupted, and long-tailed data offer indirect support for the mechanism, but this is insufficient to fully address the concern. In the revision we will add a correlation study (using the small labeled subset where ground-truth labels are available) to quantify how well pseudo-label dynamics preserve difficulty rankings. revision: yes

  2. Referee: [§4 (experimental validation)] Experiments (presumably §4) report SOTA performance on the highlighted regimes but provide no validation of pseudo-label accuracy, no ablations on labeled-subset size, and no comparison of difficulty rankings before/after pseudo-labeling. Without these, it is impossible to attribute gains to the proposed mechanism rather than incidental factors.

    Authors: We acknowledge that the submitted manuscript omits direct validation of pseudo-label accuracy, ablations over labeled-subset size, and before/after comparisons of difficulty rankings. These additions are feasible and will be included in the revised version to better isolate the contribution of the pseudo-labeling step. We will report pseudo-label accuracy on held-out labeled data, vary the labeled fraction from 1% to 10%, and show rank correlations of difficulty scores computed with versus without pseudo-labels. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method is self-contained empirical proposal

full rationale

The paper proposes SemiPrune as a practical framework that applies existing supervised pruning techniques to a pseudo-labeled pool generated via standard SSL on a small labeled subset. No equations, derivations, or parameter-fitting steps are described in the provided text that reduce a claimed prediction or result to its own inputs by construction. The central claim rests on the empirical performance of reusing off-the-shelf difficulty estimators on the SSL output rather than on any self-referential mathematical identity or load-bearing self-citation chain. This is the most common honest finding for applied method papers without internal derivations.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The method rests on the effectiveness of semi-supervised learning for generating usable pseudo-labels and the assumption that pseudo-label-induced dynamics reflect true example difficulty on the target distribution.

free parameters (1)
  • fraction of randomly labeled data
    The size of the initial labeled subset is a key design choice that affects pseudo-label quality and overall performance.
axioms (1)
  • domain assumption Semi-supervised learning produces pseudo-labels sufficiently accurate for downstream difficulty estimation on the target dataset
    This premise is required for the pseudo-labels to enable reliable coreset selection.

pith-pipeline@v0.9.0 · 5742 in / 1199 out tokens · 20883 ms · 2026-05-25T05:23:20.627483+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

65 extracted references · 65 canonical work pages · 4 internal anchors

  1. [1]

    Deep Learning Scaling is Predictable, Empirically

    Deep learning scaling is predictable, empirically , author=. arXiv preprint arXiv:1712.00409 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:1909.12673 , year=

    A constructive prediction of the generalization error across scales , author=. arXiv preprint arXiv:1909.12673 , year=

    work page arXiv 1909

  3. [3]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    Data and parameter scaling laws for neural machine translation , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    work page 2021

  4. [4]

    Advances in Neural Information Processing Systems , volume=

    Beyond neural scaling laws: beating power law scaling via data pruning , author=. Advances in Neural Information Processing Systems , volume=

    work page

  5. [5]

    LLaMA: Open and Efficient Foundation Language Models

    Llama: Open and efficient foundation language models , author=. arXiv preprint arXiv:2302.13971 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    GPT-4 Technical Report

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    International Conference on Learning Representations , year=

    An Empirical Study of Example Forgetting during Deep Neural Network Learning , author=. International Conference on Learning Representations , year=

    work page

  8. [8]

    International Conference on Learning Representations , year=

    Selection via Proxy: Efficient Data Selection for Deep Learning , author=. International Conference on Learning Representations , year=

    work page

  9. [9]

    Advances in Neural Information Processing Systems , volume=

    Identifying mislabeled data using the area under the margin ranking , author=. Advances in Neural Information Processing Systems , volume=

    work page

  10. [10]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Large-scale dataset pruning with dynamic uncertainty , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  11. [11]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Spanning training progress: Temporal dual-depth scoring (tdds) for enhanced dataset pruning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  12. [12]

    Advances in neural information processing systems , volume=

    Deep learning on a data diet: Finding important examples early in training , author=. Advances in neural information processing systems , volume=

    work page

  13. [13]

    The Eleventh International Conference on Learning Representations , year=

    Coverage-centric Coreset Selection for High Pruning Rates , author=. The Eleventh International Conference on Learning Representations , year=

    work page

  14. [14]

    Adyasha Maharana and Prateek Yadav and Mohit Bansal , booktitle=. \. 2024 , url=

    work page 2024

  15. [15]

    The Eleventh International Conference on Learning Representations , year=

    Moderate Coreset: A Universal Method of Data Selection for Real-world Data-efficient Deep Learning , author=. The Eleventh International Conference on Learning Representations , year=

    work page

  16. [16]

    Forty-first International Conference on Machine Learning , year=

    Mind the Boundary: Coreset Selection via Reconstructing the Decision Boundary , author=. Forty-first International Conference on Machine Learning , year=

    work page

  17. [17]

    Advances in neural information processing systems , volume=

    Coresets via bilevel optimization for continual learning and streaming , author=. Advances in neural information processing systems , volume=

    work page

  18. [18]

    International Conference on Machine Learning , pages=

    Probabilistic bilevel coreset selection , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    work page 2022

  19. [19]

    Forty-first International Conference on Machine Learning , year=

    Refined Coreset Selection: Towards Minimal Coreset Size under Model Performance Constraints , author=. Forty-first International Conference on Machine Learning , year=

    work page

  20. [20]

    2024 , booktitle=

    Balancing Feature Similarity and Label Variability for Optimal Size-Aware One-shot Subset Selection , author=. 2024 , booktitle=

    work page 2024

  21. [21]

    arXiv preprint arXiv:2002.03206 , year=

    Characterizing structural regularities of labeled data in overparameterized models , author=. arXiv preprint arXiv:2002.03206 , year=

    work page arXiv 2002

  22. [22]

    2023 , eprint=

    Probabilistic Bilevel Coreset Selection , author=. 2023 , eprint=

    work page 2023

  23. [23]

    International Conference on Machine Learning , pages=

    Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    work page 2025

  24. [24]

    International Conference on Machine Learning , pages=

    Grad-match: Gradient matching based data subset selection for efficient deep model training , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    work page 2021

  25. [25]

    Advances in neural information processing systems , volume=

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence , author=. Advances in neural information processing systems , volume=

    work page

  26. [26]

    Eleventh International Conference on Learning Representations , year=

    FreeMatch: Self-adaptive Thresholding for Semi-supervised Learning , author=. Eleventh International Conference on Learning Representations , year=

    work page

  27. [27]

    Bartoldson and Bhavya Kailkhura and Atul Prakash , booktitle=

    Haizhong Zheng and Elisa Tsai and Yifu Lu and Jiachen Sun and Brian R. Bartoldson and Bhavya Kailkhura and Atul Prakash , booktitle=. 2025 , url=

    work page 2025

  28. [28]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Zero-Shot Coreset Selection via Iterative Subspace Sampling , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    work page

  29. [29]

    International Conference on Learning Representations , year=

    Theoretical Analysis of Self-Training with Deep Networks on Unlabeled Data , author=. International Conference on Learning Representations , year=

    work page

  30. [30]

    Advances in neural information processing systems , volume=

    Unsupervised learning of visual features by contrasting cluster assignments , author=. Advances in neural information processing systems , volume=

    work page

  31. [31]

    arXiv preprint arXiv:2303.17896 , year=

    Exploring the limits of deep image clustering using pretrained models , author=. arXiv preprint arXiv:2303.17896 , year=

    work page arXiv

  32. [32]

    Transactions on Machine Learning Research , issn=

    Maxime Oquab and Timoth. Transactions on Machine Learning Research , issn=. 2024 , url=

    work page 2024

  33. [33]

    2026 , url=

    Effective Data Pruning through Score Extrapolation , author=. 2026 , url=

    work page 2026

  34. [34]

    Revisiting Semi-Supervised Learning in the Era of Foundation Models , author=

    work page

  35. [35]

    Advances in neural information processing systems , volume=

    Retrieve: Coreset selection for efficient and robust semi-supervised learning , author=. Advances in neural information processing systems , volume=

    work page

  36. [36]

    International Conference on Learning Representations , year=

    Deep Batch Active Learning by Diverse, Uncertain Gradient Lower Bounds , author=. International Conference on Learning Representations , year=

    work page

  37. [37]

    International Conference on Machine Learning , pages=

    Active Learning on a Budget: Opposite Strategies Suit High and Low Budgets , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    work page 2022

  38. [38]

    The Twelfth International Conference on Learning Representations , year=

    InfoBatch: Lossless Training Speed Up by Unbiased Dynamic Data Pruning , author=. The Twelfth International Conference on Learning Representations , year=

    work page

  39. [39]

    The Thirteenth International Conference on Learning Representations , year=

    Instance-dependent Early Stopping , author=. The Thirteenth International Conference on Learning Representations , year=

    work page

  40. [40]

    The Twelfth International Conference on Learning Representations , year=

    Repeated Random Sampling for Minimizing the Time-to-Accuracy of Learning , author=. The Twelfth International Conference on Learning Representations , year=

    work page

  41. [41]

    Workshop on challenges in representation learning, ICML , volume=

    Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks , author =. Workshop on challenges in representation learning, ICML , volume=. 2013 , organization=

    work page 2013

  42. [42]

    Advances in neural information processing systems , volume=

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results , author=. Advances in neural information processing systems , volume=

    work page

  43. [43]

    Advances in neural information processing systems , volume=

    Unsupervised data augmentation for consistency training , author=. Advances in neural information processing systems , volume=

    work page

  44. [44]

    Advances in neural information processing systems , volume=

    Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling , author=. Advances in neural information processing systems , volume=

    work page

  45. [45]

    The Eleventh International Conference on Learning Representations , year=

    SoftMatch: Addressing the Quantity-Quality Tradeoff in Semi-supervised Learning , author=. The Eleventh International Conference on Learning Representations , year=

    work page

  46. [46]

    Agarwalla, Abhinav , title =

    work page

  47. [47]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Revisiting Semi-Supervised Learning in the Era of Foundation Models , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    work page

  48. [48]

    International conference on machine learning , pages=

    A simple framework for contrastive learning of visual representations , author=. International conference on machine learning , pages=. 2020 , organization=

    work page 2020

  49. [49]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Momentum contrast for unsupervised visual representation learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  50. [50]

    International Conference on Machine Learning , pages=

    BWS: Best Window Selection Based on Sample Scores for Data Pruning across Broad Ranges , author=. International Conference on Machine Learning , pages=. 2024 , organization=

    work page 2024

  51. [51]

    Proceedings of the European conference on computer vision (ECCV) , pages=

    Deep clustering for unsupervised learning of visual features , author=. Proceedings of the European conference on computer vision (ECCV) , pages=

    work page

  52. [52]

    Advances in neural information processing systems , volume=

    Mixmatch: A holistic approach to semi-supervised learning , author=. Advances in neural information processing systems , volume=

    work page

  53. [53]

    Advances in neural information processing systems , volume=

    Semi-supervised learning by entropy minimization , author=. Advances in neural information processing systems , volume=

    work page

  54. [54]

    International Conference on Learning Representations , year=

    Self-labelling via simultaneous clustering and representation learning , author=. International Conference on Learning Representations , year=

    work page

  55. [55]

    European conference on computer vision , pages=

    Scan: Learning to classify images without labels , author=. European conference on computer vision , pages=. 2020 , organization=

    work page 2020

  56. [56]

    Advances in neural information processing systems , volume=

    Bootstrap your own latent-a new approach to self-supervised learning , author=. Advances in neural information processing systems , volume=

    work page

  57. [57]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Emerging properties in self-supervised vision transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

    work page

  58. [58]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Semantic-enhanced image clustering , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

    work page

  59. [59]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    work page 2021

  60. [60]

    DINOv2: Learning Robust Visual Features without Supervision

    Dinov2: Learning robust visual features without supervision , author=. arXiv preprint arXiv:2304.07193 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  61. [61]

    European conference on computer vision , pages=

    Food-101--mining discriminative components with random forests , author=. European conference on computer vision , pages=. 2014 , organization=

    work page 2014

  62. [62]

    2010 IEEE computer society conference on computer vision and pattern recognition , pages=

    Sun database: Large-scale scene recognition from abbey to zoo , author=. 2010 IEEE computer society conference on computer vision and pattern recognition , pages=. 2010 , organization=

    work page 2010

  63. [63]

    2004 conference on computer vision and pattern recognition workshop , pages=

    Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories , author=. 2004 conference on computer vision and pattern recognition workshop , pages=. 2004 , organization=

    work page 2004

  64. [64]

    2009 , url =

    Alex Krizhevsky , title =. 2009 , url =

    work page 2009

  65. [65]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

    work page 2009