Pith. sign in

REVIEW 4 major objections 6 minor 44 references

Info-Coevolution: An Efficient Framework for Data Model Coevolution

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A selective-annotation framework that fuses model and neighbor predictions cuts ImageNet labeling cost by 32% with no accuracy loss and sets its own stopping ratio.

desk verdict The selective annotation method is a real step forward, but the automatic stopping claim is unverifiable as written and needs a specified rule and seeded runs before I'd trust the 68% figure. read the letter →

arxiv 2506.08070 v2 pith:XKB4YQTE submitted 2025-06-09 cs.LG cs.AI

Ziheng Qin , Hailun Xu , Wei Chee Yew , Qi Jia , Yang Luo , Kanchan Sarkar , Danhui Guan , Kai Wang
show 1 more author
Yang You
This is my paper
classification cs.LGcs.AI
keywords dataselectionselectiveannotationinformationgainBayesianpredictionfusionactivelearningsemi-supervisedapproximatenearestneighborefficiency
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

Info-Coevolution argues that a continuously growing data stream can be curated on the fly: instead of annotating and training on everything, a model can estimate how much information a new sample would add, label only the high-gain ones, and let the model and dataset grow together. On ImageNet-1K it reports lossless accuracy (85.6% vs. 85.5% baseline) using only 68% of the annotations, with the stop ratio chosen automatically rather than tuned, and it claims the same framework selects effective unlabeled web data for semi-supervised training, halving the labeled budget to 50% at near-lossless accuracy. The method is designed to be online: it uses an approximate-nearest-neighbor index so gains are computed in near-logarithmic time per sample, avoiding the retrain-reinfer loops of active learning. If the claims hold, data-centric ML pipelines could save roughly a third of annotation cost on million-scale benchmarks without hand-tuning a selection ratio.

What carries the argument

The load-bearing object is a Bayesian Prediction Fusion of two views of a sample: the model's softmax prediction and a data-based prediction computed by confidence-weighted interpolation over the sample's nearest neighbors in feature space (Eq.4). The two views are merged via Bayesian formulas (Eqs.5-6) that update a confidence value α, and the annotation gain of a candidate is the difference between the annotator's confidence and the fused confidence. This gain drives batch selection, a dynamic-rechecking step that updates gains of neighbors after each annotation, and the automatic stopping criterion. The framework is built on a locality theorem (Theorem 3.1) asserting that nearby features yield nearby predictions, which the paper uses to justify interpolation.

What would settle it

On a small benchmark such as CIFAR-10, train a model on a 1% subset, use Eq.(4) to estimate the annotation gain of every remaining sample, and measure the rank correlation between the estimated gain and the actual accuracy improvement of adding each sample individually to the training set; if the correlation is not significantly positive, the locality/interpolation assumption that carries the method is falsified.

Watch

Extended reading notes

Core claim

The paper proposes Info-Coevolution, an online selective-annotation framework that fuses a model's prediction with a 'data view' derived from nearby already-annotated samples to estimate each candidate sample's information gain, then annotates the batch with the highest gain and dynamically rechecks neighbors. Its central empirical claim is that on ImageNet-1K this selection reaches full-data accuracy with 68% of the labels under continual supervised training, and that the same gain estimation provides an automatic stopping rule (the algorithm stops when marginal gains plateau, yielding the 68% ratio itself). It also claims compatibility with semi-supervised learning (50% labels, near-lossless) and that retrieved unlabeled web data selected by the method matches the benefit of using all such data at half the volume.

Load-bearing premise

The method assumes that within a small feature-space distance, prediction vectors of neighboring samples can be interpolated and combined to give a valid estimate of a new sample's information gain, and the paper's proof establishes only that nearby features produce nearby predictions, not that such interpolation correctly estimates entropy reduction.

Editorial extensions

If this is right

  • If the ImageNet claim holds, large-scale supervised dataset construction can cut labeling cost by roughly one third with no accuracy loss, and the saving ratio is discovered automatically rather than tuned.
  • The automatic stopping rule removes the need to annotate a holdout set to verify saturation, so data-collection pipelines can stop exactly when marginal information is exhausted.
  • Compatibility with semi-supervised training means the framework can double as a labeled-data selection front-end for SSL methods, reducing the labeled budget to 50% with near-lossless performance.
  • As an online method with ANN-based retrieval, it can be applied to continuously growing data streams, unlike batch active-learning pipelines that retrain and re-infer between rounds.
  • Retrieval-based dataset enhancement with unlabeled open-source data means the method can also curate web data for downstream tasks, not just decide which in-domain samples to label.

Reading between the lines

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

  • Editorial inference: the locality assumption suggests the method will be most effective on feature spaces that are smooth and well-behaved; tasks where decision boundaries are locally non-linear (e.g., fine-grained recognition with subtle distinctions) may see estimated gains diverge from true annotation value.
  • Editorial inference: the Bayesian fusion formulas assume independent predictors, but the model and KNN predictions are correlated since the model generates the features; if so, fused confidence may be overconfident and the reported saving ratios could partly reflect confidence calibration rather than pure information gain.
  • Editorial inference: the automatic stop rule is a direct consequence of the gain estimate; if the estimate is biased low late in training, stopping could occur before full-data performance is reached on other datasets, so users should validate the rule on a small held-out set unless the correlation between estimated and true gain is established.
  • Editorial inference: the framework's value could be tested as a drop-in module for language or multimodal data streams, where the same feature-space locality argument is plausible but unverified.
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. This paper proposes Info-Coevolution, an online selective-annotation framework. It estimates a sample's information gain by combining a model's prediction with a KNN data-view prediction via Bayesian fusion (Eqs. 4-6), rechecks neighbors dynamically after annotation, and uses the estimated gain both for prioritized selection and for automatic stopping. On ImageNet-1K the authors report lossless accuracy (85.6%) with 68% of labels under continual supervised training and 85.5% with 50% labels under Semi-ViT; on CIFAR-10 with FixMatch they report 95.85% with 1000 labels. They also explore retrieval-based enhancement with unlabeled web data and report a 1M-sample selected subset matching the performance of a 2M-sample random subset.

Significance. If the claims hold, the framework would be a practically valuable contribution: it offers annotation savings at scale, low selection overhead via approximate nearest-neighbor search, compatibility with semi-supervised learning, and a selection signal that is not purely model-uncertainty based. The public code release and the use of large-scale benchmarks are strengths. However, the strongest claim---automatic stopping without tuning---is currently under-specified, and the theoretical grounding for the fusion and interpolation steps is thinner than the paper suggests. The experimental core (ImageNet at 68%, CIFAR-10 at 1000 labels) supports the potential of the selection heuristic but does not yet verify the auto-stopping mechanism.

major comments (4)
  1. [§3.5, §4.2] The central claim of an automatic, tuning-free saving ratio is not verifiable as written. Section 3.5 defines annotation gain as cann - cmerged but neither the value of cann nor the stopping condition is specified; Section 4.2 states only that an 'automatic stop criterion' triggers when the 'marginal gain of annotating more samples is negligible' and that the 68% ratio is 'given by the algorithm itself.' Without the stopping rule (e.g., gain ≤ 0 versus gain ≤ ε for some ε), the reported 68% cannot be distinguished from a post hoc threshold choice. Please specify the stopping criterion, the cann value used, and an analysis showing that the reported ratio is insensitive to any remaining free parameter.
  2. [§3.3, Eq. (5)] Equation (5) assumes two independent predictors with confidences α1 and α2. The model view and the KNN data view are not independent: the model is trained on samples that overlap or correlate with the KNN neighbors, and both views are computed from the same feature representation. This likely makes cmerged overconfident, which would make cann - cmerged non-positive earlier than it should and could produce an artificially low stopping ratio. Please justify the independence assumption or provide an empirical calibration check (e.g., compare fused confidence with actual accuracy on a hold-out set).
  3. [§3.1, Theorem 3.1] Theorem 3.1 establishes only Lipschitz continuity of the classifier, i.e., that nearby features yield nearby predictions. The paper then states that 'this theorem implies a kind of linearity' and uses Eq. (4) to interpolate and combine logits/predictions inside an epsilon-neighborhood. Continuity does not imply that interpolation of prediction vectors or the confidence-weighted KNN fusion estimates the entropy reduction H(x) - H(x|z). This is an additional assumption that underpins both the selection scores and the stopping signal. Please either prove a stronger statement, validate the interpolation empirically, or explicitly mark Eq. (4) as a heuristic with a sensitivity study.
  4. [§4.2, Tables 1 and 3] The lossless-performance claim rests on single runs without error bars or multiple seeds. The ImageNet result (85.6% at 68% labels versus 85.6% full) and the CIFAR10 FixMatch result (95.85% with 1000 labels versus 95.85% full) are consistent with the claim, but a single run cannot establish 'lossless' on noisy deep-learning training. Additionally, the paper motivates the framework by the limitations of active learning but reports no comparison against standard active-learning baselines (e.g., uncertainty sampling, BADGE, core-set). Please add variance estimates and at least one active-learning comparison on CIFAR-10 or ImageNet to substantiate the efficiency claim.
minor comments (6)
  1. [Table 3] The method name appears as 'Info-Evolution' in the first row of Table 3; this should be 'Info-Coevolution'.
  2. [§3.2, Eq. (4)] The notation yx is described as 'the probability prediction vector of x's annotation' and Sim(x,z) is not defined; please define both and state whether Sim is cosine similarity after feature normalization.
  3. [§3.5] The batch sampling procedure is described only as 'sample with probability proportional to the gain and drop redundant samples'; the redundancy-dropping rule is not specified and is needed for reproducibility.
  4. [§3.6] The cosine retrieval threshold of 0.2 is stated without justification or sensitivity analysis; please report how this threshold was chosen and whether the results are sensitive to it.
  5. [§4.1, Appendix B] The implementation details defer to the appendix, but the appendix gives only training hyperparameters; HNSW parameters, the number of neighbors k, the distance threshold epsilon, and the model-update schedule used in Fig. 8 are missing.
  6. [§3.4, Eq. (7)] The notation H(αann) in Eq. (7) is inconsistent with the later use of cann; please clarify the relationship between confidence and entropy in this formula.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the selection-gain estimator is an explicit proxy validated against external ImageNet accuracy, and the self-citations are non-load-bearing implementation details.

full rationale

We walked the derivation chain and found no step that reduces to its own inputs by construction. The core sample-selection score is an explicit proxy: Eq. (2)-(4) estimate information gain from locality and confidence-weighted nearest-neighbor interpolation, Eq. (5)-(6) are Bayesian fusion formulas derived from an explicit independence assumption, and Eq. (7) defines annotation gain via a difference of confidences. None of these identities uses the target result (e.g., ImageNet accuracy at 68% annotation) as an input; the lossless claim is established by external benchmarking in Tab. 1 and Fig. 5, not read off from the gain estimator. The self-citations in the paper (Qin et al., 2024 for ANN de-redundancy in the superset construction; Zhou et al., 2023 as a coreset baseline) are implementation details or comparison baselines, not load-bearing premises that force the central claim. The 'automatic saving ratio without tuning' statement is under-specified: the paper does not state the stopping threshold (epsilon for 'negligible' gain) or the exact value of cann in Section 3.5, which is a reproducibility/correctness concern rather than a circular reduction. The independence assumption behind Eq. (5) is an explicit modeling assumption, and the appendix's Lipschitz proof supports only a similarity guarantee, not the full logit-interpolation claim; both are potential validity weaknesses, not circularity. No fitted parameter is renamed as a prediction, and no cited uniqueness theorem is imported from the authors' prior work. Therefore the appropriate finding is no significant circularity.

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

The framework's central gain estimate is built on a small set of assumptions: Lipschitz regularity of the classifier, linearity/interpolability of predictions in feature neighborhoods, independence of model and data views in Bayesian fusion, and a uniform or IID target distribution. These are stated in Section 3 but the linearity assumption is explicitly called an estimation choice. The free parameters (epsilon, k, annotator confidence, retrieval threshold, and the unspecified stopping threshold) are not all reported, which limits the extent to which the method can be assessed independently.

free parameters (5)
  • distance threshold epsilon = not specified
    Limits the neighborhood over which information gain is integrated in Eq.(2); inherited from Theorem 3.1 but never given a value or sensitivity analysis.
  • number of neighbors k = not specified
    Used in Eq.(4) and dynamic rechecking; k is never reported in the main text.
  • annotator confidence cann = not specified (assumed known)
    The annotation gain in Section 3.5 is computed as cann - cmerged, requiring the average annotator accuracy as an input; its value or source is not given.
  • cosine retrieval threshold = 0.2
    In Section 3.6, retrieved web samples with cosine distance larger than 0.2 are filtered; chosen without sensitivity analysis.
  • stopping threshold = not specified
    The automatic stop when marginal gain is negligible is claimed in Section 4.2 but the threshold and how it is set are not described.
assumptions (5)
  • standard math The classifier g is Lg-Lipschitz on the feature space and softmax is 1-Lipschitz.
    Used in the proof of Theorem 3.1 in Appendix A.
  • ad hoc to paper Within an epsilon-neighborhood in feature space, linearity holds enough that prediction vectors can be interpolated and combined via Eq.(4).
    The paper states 'the interpolation itself is an estimation based on linearity assumption' in Section 3.1; no empirical validation for real feature spaces.
  • domain assumption The model-view and data-view predictors are independent in the Bayesian fusion of Section 3.3.
    Eq.(5) and (6) assume two independent predictors with confidences alpha1 and alpha2; no justification for independence.
  • domain assumption The target distribution rho is uniform or IID to training data in Eq.(2)-(3).
    Stated in Section 3.1; questionable for web-sourced or streaming data.
  • domain assumption The matching distribution over wrong labels is uniform over c-1 classes when deriving lower bounds in Section 3.3.
    Used to simplify P(match with other label) = (1-alpha1)(1-alpha2)/(c-1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Info-Coevolution: An Efficient Framework for Data Model Coevolution." pith.science (2026). https://pith.science/paper/XKB4YQTE

@misc{pith2026250608070,
  author       = {Pith},
  title        = {Pith review of: Info-Coevolution: An Efficient Framework for Data Model Coevolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XKB4YQTE}},
  note         = {Machine review of arXiv:2506.08070}
}
read the original abstract

Machine learning relies heavily on data, yet the continuous growth of real-world data poses challenges for efficient dataset construction and training. A fundamental yet unsolved question is: given our current model and data, does a new data (sample/batch) need annotation/learning? Conventional approaches retain all available data, leading to non-optimal data and training efficiency. Active learning aims to reduce data redundancy by selecting a subset of samples to annotate, while it increases pipeline complexity and introduces bias. In this work, we propose Info-Coevolution, a novel framework that efficiently enables models and data to coevolve through online selective annotation with no bias. Leveraging task-specific models (and open-source models), it selectively annotates and integrates online and web data to improve datasets efficiently. For real-world datasets like ImageNet-1K, Info-Coevolution reduces annotation and training costs by 32\% without performance loss. It is able to automatically give the saving ratio without tuning the ratio. It can further reduce the annotation ratio to 50\% with semi-supervised learning. We also explore retrieval-based dataset enhancement using unlabeled open-source data. Code is available at https://github.com/NUS-HPC-AI-Lab/Info-Coevolution/.

Figures

Figures reproduced from arXiv: 2506.08070 by the authors.

Figure 1
Figure 1. The data scaling curve of different types of data. Supervised data has higher data efficiency than weak supervised data on specific downstream tasks (e.g. ImageNet), while incurring a higher collection and annotation cost. As shown in Fig.1, regarding the task-specific/downstream data efficiency(performance against data amount) and data construction cost, supervised data > weakly supervised data > unsupervised/self-… view at source ↗
Figure 2
Figure 2. A comparison of our method’s difference to active learn￾ing. Generally, active learning loops over 1. selecting samples, 2. annotating, 3. updating the model, and selecting samples again (with the updated model). In contrast, our method doesn’t have to update the model frequently. We can update the model optionally with continual supervised learning at a much lower frequency (so the training cost is also low). And o… view at source ↗
Figure 3
Figure 3. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The idea of dynamic rechecking. When we get new annotation(s), we can update the sample estimation within the ϵ-distance to better reflect the gain. It will automatically balance classes and sample density. This step is efficient with klog(n) time. Dynamic Rechecking A…
Figure 6
Figure 6. Figure 6: Compare with corset-selection SOTA method Dataset Quantization on IN1K. unlabeled data. Generalization and Robustness Theoretically, the effec￾tiveness of Info-Coevolution is model-agnostic and dataset￾agnostic. We here verify its generalization and robust￾ness across …
Figure 7
Figure 7. Figure 7: Info-Coevolution consistently improve the annotation efficiency across different datasets with both supervised/semi-supervised setting [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Ablation of model update frequency. When updating the model more frequently during selection, it shows improved data efficiency. a better annotation set at the annotation ratio. When directly selecting 10% (0.128M) annotations for ImageNet-1k with model trained on 1% d…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 13 canonical work pages

  1. [1]

    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 format.date year duplicate empty "emp...

  2. [2]

    Food-101 -- mining discriminative components with random forests

    Bossard, L., Guillaumin, M., and Van Gool, L. Food-101 -- mining discriminative components with random forests. In Fleet, D., Pajdla, T., Schiele, B., and Tuytelaars, T. (eds.), Computer Vision -- ECCV 2014, pp.\ 446--461, Cham, 2014. Springer International Publishing. ISBN 978-3-319-10599-4

    work page 2014

  3. [3]

    Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, ...

    arXiv 2020

  4. [4]

    Semi-supervised vision transformers at scale, 2022

    Cai, Z., Ravichandran, A., Favaro, P., Wang, M., Modolo, D., Bhotika, R., Tu, Z., and Soatto, S. Semi-supervised vision transformers at scale, 2022. URL https://arxiv.org/abs/2208.05688

    arXiv 2022

  5. [5]

    Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts, 2021

    Changpinyo, S., Sharma, P., Ding, N., and Soricut, R. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts, 2021. URL https://arxiv.org/abs/2102.08981

    arXiv 2021

  6. [6]

    Chen, X., Fang, H., Lin, T.-Y., Vedantam, R., Gupta, S., Dollar, P., and Zitnick, C. L. Microsoft coco captions: Data collection and evaluation server, 2015. URL https://arxiv.org/abs/1504.00325

    arXiv 2015

  7. [7]

    Selection via proxy: Efficient data selection for deep learning

    Coleman, C., Yeh, C., Mussmann, S., Mirzasoleiman, B., Bailis, P., Liang, P., Leskovec, J., and Zaharia, M. Selection via proxy: Efficient data selection for deep learning. arXiv preprint arXiv:1906.11829, 2019

    arXiv 1906

  8. [8]

    Imagenet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 248--255, 2009. doi:10.1109/CVPR.2009.5206848

    arXiv 2009

Show all 44 references
  1. [9]

    The mnist database of handwritten digit images for machine learning research [best of the web]

    Deng, L. The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE Signal Processing Magazine, 29 0 (6): 0 141--142, 2012. doi:10.1109/MSP.2012.2211477

    2012

  2. [10]

    Bert: Pre-training of deep bidirectional transformers for language understanding, 2019

    Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. URL https://arxiv.org/abs/1810.04805

    2019 arXiv

  3. [11]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. URL https://arxiv.org/ab...

    2021 arXiv

  4. [12]

    and Precioso, F

    Ducoffe, M. and Precioso, F. Adversarial active learning for deep networks: a margin based approach. arXiv preprint arXiv:1802.09841, 2018

    2018 arXiv

  5. [13]

    Deepcore: A comprehensive library for coreset selection in deep learning, 2022

    Guo, C., Zhao, B., and Bai, Y. Deepcore: A comprehensive library for coreset selection in deep learning, 2022. URL https://arxiv.org/abs/2204.08499

    2022 arXiv

  6. [14]

    Masked autoencoders are scalable vision learners, 2021

    He, K., Chen, X., Xie, S., Li, Y., Dollár, P., and Girshick, R. Masked autoencoders are scalable vision learners, 2021. URL https://arxiv.org/abs/2111.06377

    2021 arXiv

  7. [15]

    Active learning: Problem settings and recent developments

    Hino, H. Active learning: Problem settings and recent developments. CoRR, abs/2012.04225, 2020. URL https://arxiv.org/abs/2012.04225

    2012 arXiv

  8. [16]

    Submodular combinatorial information measures with applications in machine learning

    Iyer, R., Khargoankar, N., Bilmes, J., and Asanani, H. Submodular combinatorial information measures with applications in machine learning. In Algorithmic Learning Theory, pp.\ 722--754. PMLR, 2021

    2021

  9. [17]

    Grad-match: Gradient matching based data subset selection for efficient deep model training

    Killamsetty, K., Durga, S., Ramakrishnan, G., De, A., and Iyer, R. Grad-match: Gradient matching based data subset selection for efficient deep model training. In International Conference on Machine Learning, pp.\ 5464--5474. PMLR, 2021 a

    2021

  10. [18]

    Glister: Generalization based data subset selection for efficient and robust learning

    Killamsetty, K., Sivasubramanian, D., Ramakrishnan, G., and Iyer, R. Glister: Generalization based data subset selection for efficient and robust learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35-9, pp.\ 8110--8118, 2021 b

    2021

  11. [19]

    C., Lo, W.-Y., Dollár, P., and Girshick, R

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W.-Y., Dollár, P., and Girshick, R. Segment anything, 2023. URL https://arxiv.org/abs/2304.02643

    2023 arXiv

  12. [20]

    A., Bernstein, M

    Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.-J., Shamma, D. A., Bernstein, M. S., and Li, F.-F. Visual genome: Connecting language and vision using crowdsourced dense image annotations, 2016. URL https://arxiv.org/abs/1602.07332

    2016 arXiv

  13. [21]

    Cifar-10 (canadian institute for advanced research)

    Krizhevsky, A., Nair, V., and Hinton, G. Cifar-10 (canadian institute for advanced research). a . URL http://www.cs.toronto.edu/ kriz/cifar.html

  14. [22]

    Cifar-100 (canadian institute for advanced research)

    Krizhevsky, A., Nair, V., and Hinton, G. Cifar-100 (canadian institute for advanced research). b . URL http://www.cs.toronto.edu/ kriz/cifar.html

  15. [23]

    A survey on deep active learning: Recent advances and new frontiers, 2024

    Li, D., Wang, Z., Chen, Y., Jiang, R., Ding, W., and Okumura, M. A survey on deep active learning: Recent advances and new frontiers, 2024. URL https://arxiv.org/abs/2405.00334

    2024 arXiv

  16. [24]

    BLIP : Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Li, J., Li, D., Xiong, C., and Hoi, S. BLIP : Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., and Sabato, S. (eds.), Proceedings of the 39th International Conf...

    2022

  17. [25]

    Malkov, Y. A. and Yashunin, D. A. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs, 2018. URL https://arxiv.org/abs/1603.09320

    2018 arXiv

  18. [26]

    Active learning by acquiring contrastive examples

    Margatina, K., Vernikos, G., Barrault, L., and Aletras, N. Active learning by acquiring contrastive examples. arXiv preprint arXiv:2109.03764, 2021

    2021 arXiv

  19. [27]

    Coresets for data-efficient training of machine learning models

    Mirzasoleiman, B., Bilmes, J., and Leskovec, J. Coresets for data-efficient training of machine learning models. In International Conference on Machine Learning, pp.\ 6950--6960. PMLR, 2020

    2020

  20. [28]

    Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., and Ng, A. Y. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. URL http://ufldl.stanford.edu/housenumbers/nips2011_housen...

    2011

  21. [29]

    Im2text: Describing images using 1 million captioned photographs

    Ordonez, V., Kulkarni, G., and Berg, T. Im2text: Describing images using 1 million captioned photographs. In Shawe-Taylor, J., Zemel, R., Bartlett, P., Pereira, F., and Weinberger, K. (eds.), Advances in Neural Information Processing Systems, volume 24. Curran Associates, Inc....

    2011

  22. [30]

    Paul, M., Ganguli, S., and Dziugaite, G. K. Deep learning on a data diet: Finding important examples early in training. Advances in neural information processing systems, 34: 0 20596--20607, 2021

    2021

  23. [31]

    Dataset growth

    Qin, Z., Xu, Z., Zhou, Y., Zheng, Z., Cheng, Z., Tang, H., Shang, L., Sun, B., Peng, X., Timofte, R., Yao, H., Wang, K., and You, Y. Dataset growth. In Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part IX, pp.\...

    2024 doi

  24. [32]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning transferable visual models from natural language supervision, 2021. URL https://arxiv.org/abs/2103.00020

    2021 arXiv

  25. [33]

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., and Liu, P. J. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21 0 (140): 0 1--67, 2020

    2020

  26. [34]

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs, 2021

    Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., and Komatsuzaki, A. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs, 2021. URL https://arxiv.org/abs/2111.02114

    2021 arXiv

  27. [36]

    and Savarese, S

    Sener, O. and Savarese, S. Active learning for convolutional neural networks: A core-set approach, 2018. URL https://arxiv.org/abs/1708.00489

    2018 arXiv

  28. [37]

    Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning

    Sharma, P., Ding, N., Goodman, S., and Soricut, R. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018

    2018

  29. [38]

    Small-gan: Speeding up gan training using core-sets

    Sinha, S., Zhang, H., Goyal, A., Bengio, Y., Larochelle, H., and Odena, A. Small-gan: Speeding up gan training using core-sets. In International Conference on Machine Learning, pp.\ 9005--9015. PMLR, 2020

    2020

  30. [39]

    B., Kirsch, A., Farquhar, S., Gal, Y., Foster, A., and Rainforth, T

    Smith, F. B., Kirsch, A., Farquhar, S., Gal, Y., Foster, A., and Rainforth, T. Prediction-oriented bayesian active learning, 2023. URL https://arxiv.org/abs/2304.08151

    2023 arXiv

  31. [40]

    D., Kurakin, A., Zhang, H., and Raffel, C

    Sohn, K., Berthelot, D., Li, C.-L., Zhang, Z., Carlini, N., Cubuk, E. D., Kurakin, A., Zhang, H., and Raffel, C. Fixmatch: Simplifying semi-supervised learning with consistency and confidence, 2020. URL https://arxiv.org/abs/2001.07685

    2020 arXiv

  32. [41]

    Toneva, M., Sordoni, A., Combes, R. T. d., Trischler, A., Bengio, Y., and Gordon, G. J. An empirical study of example forgetting during deep neural network learning. arXiv preprint arXiv:1812.05159, 2018

    2018 arXiv

  33. [42]

    Freematch: Self-adaptive thresholding for semi-supervised learning, 2023

    Wang, Y., Chen, H., Heng, Q., Hou, W., Fan, Y., Wu, Z., Wang, J., Savvides, M., Shinozaki, T., Raj, B., Schiele, B., and Xie, X. Freematch: Self-adaptive thresholding for semi-supervised learning, 2023. URL https://arxiv.org/abs/2205.07246

    2023 arXiv

  34. [43]

    Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling, 2022

    Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., and Shinozaki, T. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling, 2022. URL https://arxiv.org/abs/2110.08263

    2022 arXiv

  35. [44]

    Dataset quantization, 2023

    Zhou, D., Wang, K., Gu, J., Peng, X., Lian, D., Zhang, Y., You, Y., and Feng, J. Dataset quantization, 2023. URL https://arxiv.org/abs/2308.10524

    2023 arXiv

  36. [45]

    Aligning books and movies: Towards story-like visual explanations by watching movies and reading books

    Zhu, Y. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. arXiv preprint arXiv:1506.06724, 2015

    2015 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.