Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

DIET-CP: Lightweight and Data Efficient Self Supervised Continued Pretraining

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read DIET-CP: steering a released vision backbone to a new domain with 1000 unlabeled images and no extra hyperparameters.

desk verdict DIET-CP's core recipe is a genuinely cheap and effective way to steer released ViTs toward out-of-domain data, but the abstract oversells its stability and hyperparameter-freedom. read the letter →

arxiv 2509.06990 v1 pith:LTLHZKW3 submitted 2025-09-02 cs.CV cs.LG

Bryan Rodas , Natalie Montesino , Jakob Ambsdorf , David Klindt , Randall Balestriero This is my paper
classification cs.CVcs.LG
keywords continuedpretrainingself-supervisedlearninginstancediscriminationfoundationmodelstransferlabelefficiencymedicalimageclassificationvisiontransformers
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 proposes DIET-CP, a continued-pretraining recipe that takes a released vision backbone (weights only) and steers it toward a small target dataset by training a linear head to predict each image's index via cross-entropy, then unfreezing the last two transformer blocks. It claims that with roughly 1000 unlabeled images and no more hyperparameters than supervised finetuning, this improves downstream k-NN and linear-probing classification, especially for out-of-domain tasks such as medical imaging and galaxy morphology. The authors report average F1 gains of 17.77 and 12.44 points on k-NN and 4.81 and 4.43 points on linear probing for DINOv2 and DINOv3 on MedMNIST datasets, while noting that in-domain fine-grained tasks can degrade. If correct, anyone with a released ViT backbone can cheaply adapt it to a specialized domain in under ten GPU-minutes without needing labels or teacher checkpoints.

What carries the argument

The DIET objective: a cross-entropy classification task where each training image is its own class, with no projector network, teacher, or positive pairs. It carries the argument by converting continued pretraining into a standard supervised-style optimization on the backbone only, and is paired with a simple recipe: freeze the backbone for the first 5% of epochs, then unfreeze the last two transformer blocks and train jointly with a linear head for 150 epochs with label smoothing near 0.3, weight decay 0.05, and cosine learning-rate decay. The paper relies on a theoretical result that index-discrimination cross-entropy recovers linearly decodable ground-truth factors of the data-generating

What would settle it

Apply DIET-CP to 1000 fixed images with randomly permuted pixel content while keeping indices fixed; if k-NN on real held-out images still improves, the gain is not target-domain steering. A second check: compare DIET-CP against supervised finetuning on the same 1000 labeled images and the same compute budget—the paper's claim of matching supervised-finetuning hyperparameters predicts DIET-CP should match or beat that baseline on out-of-domain tasks.

Watch

Extended reading notes

Core claim

The central claim is that continued pretraining with the DIET objective—cross-entropy on the datum index, L_DIET(x_n) = XEnt(W f_theta(x_n), n)—is a lightweight and data-efficient way to adapt a foundation model to a new distribution using only the released backbone weights. The paper's experiments show that this objective, applied to a random subset of about 1000 unlabeled target images, consistently improves unsupervised k-NN clustering and linear-probing classification across medical imaging datasets and on a galaxy-morphology dataset, for three different vision backbones (DINOv2, DINOv3, and MAE). The gains are largest on out-of-domain tasks, while in-domain fine-grained visual categoriz

Load-bearing premise

DIET-CP's gains depend on a small unlabeled image set being enough for index-discrimination to uncover the target domain's real structure rather than simply memorizing image-to-index mappings, and the paper relies on a theoretical result whose assumptions it does not verify on the medical or galaxy images.

Editorial extensions

If this is right

  • Released backbone weights alone are sufficient for continued pretraining; no teacher weights, optimizer state, projector, or self-distillation machinery is required.
  • Small unlabeled datasets (around 500–1000 images) are enough to produce a measurable downstream performance gain, and more samples did not yield additional benefit in the reported ablation.
  • The method improves unsupervised k-NN performance substantially more than linear probing, suggesting it primarily re-organizes the representation space into better clusters.
  • Out-of-domain tasks benefit consistently across backbones, while fine-grained in-domain tasks can degrade, so practitioners should apply DIET-CP selectively and evaluate on a held-out set.
  • The recipe uses the same hyperparameters as supervised finetuning, making label-free continued pretraining practical on a single GPU in a short wall-clock time.

Reading between the lines

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

  • Because the recipe only unfreezes the last two transformer blocks, DIET-CP likely acts more as a lightweight distribution adapter than as full re-pretraining; a direct test would be to vary the number of unfrozen blocks and measure the trade-off between target-domain gains and preservation of original features.
  • The paper's own loss curves show DIET loss decreases monotonically even when downstream performance plateaus, so the loss is not a reliable early-stopping signal; a label-free proxy such as nearest-neighbor consistency on held-out images could be tested instead.
  • The sample-count ablation suggests a sweet spot near 1000 images; an untested implication is that the point at which more samples stop helping could serve as a rough measure of the target domain's accessible structure or intrinsic dimensionality.
  • RetinaMNIST, the only ordinal-regression task, is the dataset where linear probing degrades for both DINO models; a testable extension would be to treat DIET-CP class indices as ordinal targets or add a rank-preserving head for such tasks.
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 / 5 minor

Summary. The paper proposes DIET-CP, a continued-pretraining method that adapts a released vision backbone to a target domain using only unlabeled images. The objective is cross-entropy with the datum's index as the class label (Eq. 1), applied for 150 epochs with a frozen backbone for the first 5% of epochs and then partial unfreezing of the last two transformer blocks (Appendix B). The authors evaluate on nine MedMNIST datasets with three ViT-B backbones (DINOv2, DINOv3, MAE), plus FGVC-Aircraft, Food-101, and Galaxy10-DECaLS, reporting k-NN and linear-probe F1 before/after DIET-CP. The main empirical finding is that out-of-domain tasks (medical images, galaxy morphology) improve substantially, e.g., +17.77 and +12.44 k-NN F1 for DINOv2 and DINOv3 on medical datasets, while in-domain fine-grained tasks often degrade, especially for linear probing. The paper also includes a sample-size ablation and a model-size ablation.

Significance. If the method works as advertised, it would be a practically valuable tool: adapting a released backbone to a small specialized domain in under ten GPU-minutes with no labels and no additional hyperparameters beyond those of supervised finetuning. The empirical gains on MedMNIST and Galaxy10 are large, consistent across three backbones for k-NN, and reported with standard deviations over three runs in Table 3. The method is simple enough to be easily reproduced. However, the paper's own Table 2 shows systematic linear-probe degradation on FGVC-Aircraft and Food-101 for DINOv2/DINOv3, undermining the abstract's unqualified 'stable across data modalities and backbone choices' and 'significant performance boost' claims. The 'no more hyperparameters than supervised finetuning' claim is also not backed by a supervised-finetuning baseline or ablations of the tuned recipe components (label smoothing ~0.3, 5% frozen warmup, last-two-block unfreezing). The theoretical link to [12] is plausible but its assumptions are not checked on the evaluated datasets. Overall, the paper reports a genuinely promising technique with clear out-of-domain benefits, but the current presentation overclaims and

major comments (4)
  1. [Abstract and Section 2] The claim that DIET-CP 'introduces no more hyperparameters than supervised finetuning' is unsupported. Section 2 and Appendix B specify label smoothing of ~0.3 (found after tuning), a 5%-of-epochs frozen-backbone warmup, unfreezing exactly the last two blocks, 150 epochs, learning rate 1e-4, batch size 32, weight decay 0.05, and a specific augmentation list. These are additional design choices not mandated by supervised finetuning, and no supervised-finetuning baseline is reported to demonstrate the claimed equivalence. Without ablations showing that the default recipe is insensitive to label smoothing and the block-unfreezing split, the practical out-of-the-box claim fails where the paper says it is strongest.
  2. [Table 2] The abstract states the method is 'stable across data modalities and backbone choices' and provides 'a significant performance boost,' but Table 2 shows systematic linear-probe degradation for DINOv2 and DINOv3 on in-domain fine-grained tasks: FGVC-Aircraft LP drops by 5.00 and 12.44 F1 points, and Food-101 LP drops by 8.25 and 8.60 F1 points for DINOv2 and DINOv3 respectively. Section 3 explicitly acknowledges that DIET-CP 'deteriorates ... in some cases for fine-grained in-domain tasks.' The unqualified abstract claim is therefore contradicted by the paper's own data; the claim should be restricted to out-of-domain tasks at minimum, and ideally accompanied by a predictive criterion for when the method helps.
  3. [Section 2.1 / Figure 2] The sample-size ablation (Figure 2) uses 'a constant set of 1000 labels' for training the k-NN and LP classifiers, but the main experimental protocol in Section 2.1 does not state how many labeled samples are used to fit the k-NN/LP readouts in Tables 1 and 2. Since the readout complexity affects the measured gain, the protocol should be stated precisely and consistently, and the evaluation should be clearly separated from the unlabeled DIET-CP stage. As written, a reader cannot determine whether the pre/post comparison uses the same readout setup, which is essential for interpreting the reported improvements.
  4. [Section 2 (theoretical justification)] The paper justifies the DIET objective by citing the theorem from [12] that instance discrimination recovers ground-truth factors under certain assumptions, 'provably yielding linearly decodable representations.' However, no evidence is provided that these assumptions hold for MedMNIST or Galaxy10 images, nor is the theorem's transfer to the partial-unfreezing recipe addressed. This is a support gap for the claim that index-discrimination on a small real-image set reorganizes representations along target-domain structure rather than merely memorizing indices. The empirical results are still meaningful, but the theoretical framing is stronger than the evidence.
minor comments (5)
  1. [Appendix B] Typo: 'DEIT head' should be 'DIET head'.
  2. [Figure 2] Typo: 'using a a DINOv2 ViT-S' should read 'using a DINOv2 ViT-S'.
  3. [Section 2.1] The phrase 'less then 1000 images' should be 'less than 1000 images' (also appears in Table 6 caption for BreastMNIST).
  4. [Table 2] Table 2 reports only means, while Table 1 reports standard deviations in the full table (Table 3). Adding standard deviations to the non-medical results would allow assessment of whether the degradations are significant.
  5. [Section 1 / References] The introduction says 'free of additional hyperparameters' but the method section introduces label smoothing and block-unfreezing choices; please reconcile this wording with the actual recipe.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical results are externally benchmarked and not forced by the loss definition.

full rationale

The paper's central claim is empirical: applying the DIET objective (cross-entropy on the datum index) as continued pretraining on about 1000 unlabeled target images improves held-out k-NN and linear-probing classification. The loss is not defined in terms of the evaluation metric; the reported gains are measured on validation splits of external benchmarks (MedMNIST, Galaxy10, FGVC-Aircraft, Food-101), and the paper's own failure cases (e.g., FGVC-Aircraft and Food-101 linear probing degradation, RetinaMNIST LP degradation) show the outcome is not forced by construction. The theoretical justification is cited from prior work with overlapping authors (refs. [8], [12]), but that work is a published, general result about the DIET objective, not fitted to the present benchmarks, and the present paper does not use the cited theory to forbid alternatives or to define its measurements. The 'no more hyperparameters than supervised finetuning' assertion is an unsupported claim about the recipe (label smoothing ~0.3, frozen first 5% of epochs, unfreezing last two blocks), but it is not circular: the reported improvements are not logically entailed by that claim, nor are the evaluation numbers defined by it. There is self-citation, but it is not load-bearing in the sense of making the empirical prediction equivalent to an input; the experiments stand as independent falsifiable evidence.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper contributes a recipe rather than a derivation, so its weight falls on empirical evaluation plus a self-cited theory. All practical claims depend on hand-chosen hyperparameters (label smoothing near 0.3, unfreezing the last two blocks, lr 1e-4, 150 epochs) that are fixed across datasets rather than fitted per dataset; the downstream metric itself is external and not fitted. The theoretical guarantee invoked is from the authors' own prior work and its assumptions are unverified on these datasets. No invented entities are introduced.

free parameters (4)
  • Label smoothing value = ~0.3
    Chosen by the authors as 'best in our setup' (Section 2); the paper claims no more hyperparameters than supervised finetuning, but this value was hand-tuned.
  • Learning rate and schedule = 1e-4, 150 epochs, 10% warmup, cosine annealing
    Fixed recipe shared across all runs (Appendix B); values are hand-chosen, and the method's stability claims rest on them.
  • Frozen/unfrozen block split = last two transformer blocks; 5% frozen warmup
    Recipe choice in Section 2.1 and Appendix B; Section 3 admits determining how many layers to train remains open.
  • DIET class count N (training subset size) = 1000 (546 for BreastMNIST, 800 for Galaxy10-DECaLS)
    Number of index classes equals the number of images in the random training subset (Appendix D); chosen by the authors.
assumptions (3)
  • domain assumption The theorem of Reizinger et al. [12] applies: cross-entropy instance discrimination provably recovers ground-truth factors and gives linearly decodable representations on these real imaging distributions.
    Invoked in Section 2 as the reason DIET-CP should steer the backbone; the theorem's recoverability assumptions are not checked on MedMNIST or Galaxy10 images.
  • ad hoc to paper Partially unfreezing the backbone (last two blocks only) is enough to steer the representation without destroying pretrained features.
    A recipe assumption with no ablation on block depth; the conclusion in Section 3 lists layer choice as unresolved.
  • domain assumption The linear-probe and k-NN evaluation on the validation split measures general domain alignment and is not inflated by the backbone's unsupervised exposure to the training subset.
    Evaluation is on held-out validation and probes use training-set labels, so leakage risk is limited, but probe protocols are not fully specified (probe label count, optimizer, k).

how reviews work

0 comments
Cite this review

Pith. "Pith review of DIET-CP: Lightweight and Data Efficient Self Supervised Continued Pretraining." pith.science (2026). https://pith.science/paper/LTLHZKW3

@misc{pith2026250906990,
  author       = {Pith},
  title        = {Pith review of: DIET-CP: Lightweight and Data Efficient Self Supervised Continued Pretraining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LTLHZKW3}},
  note         = {Machine review of arXiv:2509.06990}
}
read the original abstract

Continued pretraining offers a promising solution for adapting foundation models to a new target domain. However, in specialized domains, available datasets are often very small, limiting the applicability of SSL methods developed for large-scale pretraining and making hyperparameter search infeasible. In addition, pretrained models are usually released as backbone-weights only, lacking important information to continue pretraining. We propose to bridge this gap with DIET-CP, a simple continued pretraining strategy, where any strong foundation model can be steered towards the new data distribution of interest. DIET-CP relies on a very simple objective, requires no labels, and introduces no more hyperparameters than supervised finetuning. It is stable across data modalities and backbone choices, while providing a significant performance boost for state-of-the-art models such as DINOv3 using only 1000 images.

Figures

Figures reproduced from arXiv: 2509.06990 by the authors.

Figure 1
Figure 1. DIET-CP is a label-free and efficient method for steering foundation models towards a data [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Ablation study over the number of samples used for DIET-CP of a DINOv2 ViT-S. For [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. DIET uses the datum index (n) as the class-target –effectively turning unsupervised learning into a supervised learning problem. In our case, we employ the cross-entropy loss (X-Ent), no extra care needed to handle different dataset or architectures. As opposed to current SOTA, we do not rely on a projector nor positive views i.e no change needs to be done to any existing supervised pipeline to obtain DIET. Figure a… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: DIET loss curves for DINOv2 ViT-S and corresponding [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: DIET loss curves, k-NN and linear probing accuracy for ViT-B DINOv2, DINOv3, and MAE on PathMNIST. Backbones reach different loss levels, but they are not strongly correlated to downstream performance. is measured as F1 score for k-NN and linear probing and averaged ov…
Figure 6
Figure 6. Figure 6: t-SNE plot of pre and post DIET-CP representations for MAE on PathMNIST. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: t-SNE plot of pre and post DIET-CP representations for DINOv2 on PathMNIST. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics

    cs.LG 2025-11 conditional novelty 6.0 of 10

    LeJEPA derives an optimal isotropic Gaussian target for embeddings and enforces it via sketched regularization to deliver scalable, heuristics-free self-supervised pretraining with 79% ImageNet linear accuracy on ViT-H/14.

Reference graph

Works this paper leans on

29 extracted references · 19 canonical work pages · cited by 1 Pith paper

  1. [12]

    Cross-entropy is all you need to invert the data generating process

    Patrik Reizinger, Alice Bizeul, Attila Juhos, Julia E V ogt, Randall Balestriero, Wieland Brendel, and David Klindt. Cross-entropy is all you need to invert the data generating process. In The Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  2. [1]

    Dinobloom: a foundation model for generalizable cell embeddings in hematology

    Valentin Koch, Sophia J Wagner, Salome Kazeminia, Ece Sancar, Matthias Hehr, Julia A Schnabel, Tingying Peng, and Carsten Marr. Dinobloom: a foundation model for generalizable cell embeddings in hematology. In International Conference on Medical Image Computing and Computer-Assisted Interven- tion, pages 520–530. Springer, 2024

    work page 2024

  3. [2]

    General Methods Make Great Domain-specific Foundation Models: A Case-study on Fetal Ultrasound

    Jakob Ambsdorf, Asbjørn Munk, Sebastian Llambias, Anders Nymark Christensen, Kamil Mikolaj, Randall Balestriero, Martin Tolsgaard, Aasa Feragen, and Mads Nielsen. General methods make great domain-specific foundation models: A case-study on fetal ultrasound. arXiv preprint arXiv:2506.19552, 2025

    work page Pith review arXiv 2025

  4. [3]

    Continual pre-training of large language models: How to re-warm your model? In Workshop on Efficient Systems for Foundation Models@ ICML2023, 2023

    Kshitij Gupta, Benjamin Thérien, Adam Ibrahim, Mats Leon Richter, Quentin Gregory Anthony, Eugene Belilovsky, Irina Rish, and Timothée Lesort. Continual pre-training of large language models: How to re-warm your model? In Workshop on Efficient Systems for Foundation Models@ ICML2023, 2023

    work page 2023

  5. [4]

    Reuse, don’t retrain: A recipe for continued pretraining of language models

    Jupinder Parmar, Sanjev Satheesh, Mostofa Patwary, Mohammad Shoeybi, and Bryan Catanzaro. Reuse, don’t retrain: A recipe for continued pretraining of language models. arXiv preprint arXiv:2407.07263, 2024

    arXiv 2024

  6. [5]

    Efficient domain continual pretraining by mitigating the stability gap

    Yiduo Guo, Jie Fu, Huishuai Zhang, and Dongyan Zhao. Efficient domain continual pretraining by mitigating the stability gap. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar, editors, Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 32850–32870, Vienna, Aus...

    work page 2025

  7. [6]

    Oriane Siméoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julie...

    arXiv 2025

  8. [7]

    V o, et al

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V . V o, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    arXiv 2023

Show all 29 references
  1. [8]

    Occam’s razor for self supervised learning: What is sufficient to learn good representations? arXiv preprint arXiv:2406.10743, 2024

    Mark Ibrahim, David Klindt, and Randall Balestriero. Occam’s razor for self supervised learning: What is sufficient to learn good representations? arXiv preprint arXiv:2406.10743, 2024

    2024 arXiv

  2. [9]

    Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021

    Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021

    2021 arXiv

  3. [10]

    Imagenet: A large-scale hierarchical image database

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

    2009

  4. [11]

    Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study

    Jakob Nikolas Kather, Johannes Krisam, et al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLOS Medicine, 16(1):1–22, 01 2019

    2019

  5. [13]

    Image BERT pre-training with online tokenizer

    Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In International Conference on Learning Representations, 2022

    2022

  6. [14]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979–15988, 2022

    2022

  7. [15]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv...

    2010 arXiv

  8. [16]

    Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis

    Jiancheng Yang, Rui Shi, and Bingbing Ni. Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis. In IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 191–195, 2021

    2021

  9. [17]

    Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification

    Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10(1):41, 2023

    2023

  10. [18]

    Dataset of breast ultrasound images

    Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in Brief, 28:104863, 2020

    2020

  11. [19]

    The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions

    Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, page 180161, 2018

    2018

  12. [20]

    Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic)

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging ...

    2018 arXiv

  13. [21]

    Kermany, Michael Goldbaum, et al

    Daniel S. Kermany, Michael Goldbaum, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell, 172(5):1122 – 1131.e9, 2018

    2018

  14. [22]

    The liver tumor segmentation benchmark (lits)

    Patrick Bilic, Patrick Ferdinand Christ, et al. The liver tumor segmentation benchmark (lits). CoRR, abs/1901.04056, 2019

    1901 arXiv

  15. [23]

    X. Xu, F. Zhou, et al. Efficient multiple organ localization in ct image using 3d region proposal network. IEEE Transactions on Medical Imaging, 38(8):1885–1898, 2019

    2019

  16. [24]

    Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases

    Xiaosong Wang, Yifan Peng, et al. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, pages 3462–3471, 2017

    2017

  17. [25]

    Galaxy10 DECaLS dataset

    astroNN. Galaxy10 DECaLS dataset. https://astronn.readthedocs.io/en/latest/ galaxy10.html, 2019

    2019

  18. [26]

    Galaxy zoo decals: Detailed visual morphology measurements from volunteers and deep learning for 314 000 galaxies

    Mike Walmsley, Chris Lintott, Tobias Géron, Sandor Kruk, Coleman Krawczyk, Kyle W Willett, Steven Bamford, Lee S Kelvin, Lucy Fortson, Yarin Gal, et al. Galaxy zoo decals: Detailed visual morphology measurements from volunteers and deep learning for 314 000 galaxies. Monthly N...

    2022

  19. [27]

    Fine-grained visual classification of aircraft

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013

    2013 arXiv

  20. [28]

    Food-101: Mining discriminative components with random forests

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101: Mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), volume 8694 of Lecture Notes in Computer Science, pages 446–461. Springer, 2014

    2014

  21. [29]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6 A DIET sample1 sample2 Training dataset . . . sampleN select samplen N-way classifier Deep Network XEnt(n,Wfθ(samplen)) • no siamese/teache...

    2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.