Rethinking the Bias of Foundation Model under Long-tailed Distribution

Bing Su; Bin Qin; Hao Chen; Jiahao Chen; Jiangmeng Li

REVIEW 4 major objections 5 minor 40 references

Rethinking the Bias of Foundation Model under Long-tailed Distribution

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Foundation models carry hidden bias from pre-training data that ordinary re-balancing cannot fix; the paper proposes a causal backdoor adjustment that averages three models and lifts accuracy ~1.67% per dataset.

desk verdict Useful empirical decomposition of upstream vs downstream imbalance in fine-tuned foundation models, but the 'backdoor adjustment' is really an equal-weight ensemble and the causal story should be cut or substantially rewritten. read the letter →

arxiv 2501.15955 v3 pith:AFUEXSQA submitted 2025-01-27 cs.LG cs.CVstat.ML

Jiahao Chen , Bin Qin , Jiangmeng Li , Hao Chen , Bing Su This is my paper · ORCID

classification cs.LGcs.CVstat.ML

keywords long-tailedlearningfoundationmodelsparameterimbalancedatacausalinferencebackdooradjustmentlogitparameter-efficientfine-tuning

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 argues that when a foundation model is fine-tuned for a long-tailed downstream task, the model carries two distinct kinds of bias: data imbalance from the downstream training set and parameter imbalance inherited from the imbalanced pre-training data. The paper's central claim is that parameter imbalance dominates during fine-tuning and cannot be removed by standard re-balancing tricks such as logit adjustment, which only fix the classifier. To address both imbalances, the paper treats the incomplete semantic factor — the partial visual features a biased pre-trained model learns, like recognizing a dog only by its head — as a confounder, and applies a backdoor adjustment that averages predictions across multiple foundation models. The method reports consistent gains, roughly 1.67% average per dataset, with the largest improvements on tail classes.

What carries the argument

The key machinery is backdoor adjustment from causal inference. The paper models an incomplete semantic factor $C$ — a partial feature captured by a biased pre-trained model, such as focusing on a dog's head instead of its whole body — as a confounder that opens the spurious path $X\leftarrow C\rightarrow Y$. By estimating $P(Y|\operatorname{do}(x)) = \sum_c P(Y|b,c)P(c)$ with balanced representation $b$, the method removes the confounding effect; with $P(c)=1/M$ and $M=3$ models, the operation is an equal-weight fusion of three fine-tuned, logit-adjusted models.

What would settle it

A simple test construction is given by the paper's own E.9 setup: train the same model on data generated with unbalanced incomplete semantic factors (e.g. 90 percent dog-head images) and observe a large performance drop on a balanced test set, which would confirm the confounder effect; conversely, if varying the prior of the incomplete semantic factors across models produces no drop and averaging three arbitrary models gives the same gain, the causal mechanism is not needed.

Watch

Extended reading notes

Core claim

The paper finds that fine-tuning a foundation model on a long-tailed task inherits two separate biases: data imbalance from the downstream data and parameter imbalance from the pre-training data. Parameter imbalance, measured by the estimated label prior of the inaccessible pre-training data, persists after fine-tuning and is not mitigated by logit adjustment or other re-balancing losses applied during training. The proposed remedy is a backdoor adjustment that estimates $P(Y|\operatorname{do}(x))$ instead of $P(Y|x)$, treating the incomplete semantic factor $C$ as a confounder on the path $X\leftarrow C\rightarrow Y$. Concretely, the method fine-tunes three CLIP-family models (CLIP, OpenCLIP, MetaCLIP) with a re-balanced loss, treats each as an instantiation of an incomplete semantic factor, and averages their predictions under a uniform prior $P(c)=1/M$. The paper reports that this adjustment yields an average gain of about 1.67% per dataset over the strongest baselines on ImageNet-LT, Places365-LT, and iNaturalist2018.

Load-bearing premise

The entire causal interpretation rests on treating CLIP, OpenCLIP, and MetaCLIP as three draws from the confounder distribution $C$ with equal prior; if they are not exchangeable instantiations of the same incomplete semantic factor, the backdoor adjustment is just an equal-weight ensemble.

Editorial extensions

If this is right

If parameter imbalance is the dominant bias in fine-tuned foundation models, then improving tail-class performance on a downstream task requires acting on pre-trained weights, not just on the downstream loss or classifier.
Logit adjustment and similar re-balancing techniques should be re-evaluated for PEFT-based long-tailed learning: they help data imbalance by improving the classifier but leave parameter imbalance largely untouched.
Ensembling multiple foundation models that have different pre-training data biases becomes a practical way to deconfound downstream predictions, with gains that grow as more diverse models are added.
The backdoor adjustment gives a concrete way to improve fairness across head and tail classes simultaneously, without the trade-off usually seen when re-balancing only the classifier.
The causal framing suggests that any future foundation model with known different pre-training imbalance could be plugged into this adjustment without retraining.

Reading between the lines

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

Editorial inference: the uniform prior $P(c)=1/M$ makes the method an equal-weight ensemble; if the three models are not exchangeable draws from the confounder distribution, the causal language is not strictly needed to describe the result — averaging three logit-adjusted models may still work as an ensembling trick.
Editorial inference: the method could be tested against a cheaper alternative, e.g. fine-tuning one model with three different re-balancing seeds or three different PEFT adapters, to see whether the gains come from model diversity or from the causal adjustment itself.
Editorial inference: the paper's own E.4 experiments suggest that adding more tail data can mitigate parameter imbalance in the Reverse setting, implying a testable conjecture that data augmentation for tail classes could substitute for extra foundation models.
Editorial inference: because the method averages three fine-tuned models, its FLOPs in the paper grow linearly with $M$; the evaluation of trade-off between compute and accuracy gains is left open.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: this paper has a solid empirical finding and a method that works modestly well, but the causal packaging doesn't survive contact with the equations. The parameter-imbalance vs data-imbalance distinction is genuinely useful, and the negative result that GLA-Train fails to fix parameter imbalance while LA fixes data imbalance is a real contribution. The proposed method—fine-tune CLIP, OpenCLIP, and MetaCLIP with logit adjustment and average their predictions—is simple and consistently beats single-model LA by roughly 1.5-3% on the three benchmarks. That part is reproducible and worth knowing.

The soft spot is the claim that this is a backdoor adjustment that 'learns the true causal effect.' Eq. (7) reduces to a uniform average over three LA-adjusted models. For it to be backdoor adjustment, C has to be a well-defined confounder with P(c)=1/M justified, and CLIP/OpenCLIP/MetaCLIP have to be instantiations or draws from that distribution. None of that is established. The paper defines C via the models, then uses the models to estimate P(Y|b,c), which is circular. The Grad-CAM and synthetic experiments in the appendix show that different models attend to different parts and that a known confounder can hurt, but they don't show these three checkpoints are exchangeable draws from a common C. So the causal interpretation is unsupported. The practical method survives—it's an ensemble of three LA-debiased classifiers, and the M-ablation shows adding models helps, which is exactly what you'd expect from ensembling. But the paper overclaims by calling it causal.

Two more issues: the estimated pretraining prior from GLA is load-bearing for the parameter-imbalance analysis and is fitted on the validation set without verification, and the main tables have no error bars. These are addressable. The missing baseline—plain average of the three LA models' logits without any causal framing—is the one that would let a reader decide whether the causal story adds anything beyond ensembling. Citation pattern is fine; they build directly on GLA and LA.

Who this is for: practitioners fine-tuning CLIP-like models on long-tailed data will get a useful takeaway and a cheap method. The paper deserves a serious referee, but a referee should push hard on the causal claim and the missing ensemble baseline. If it's resubmitted, I'd want the causal language cut back to 'we ensemble diverse LA-fine-tuned models' unless identification is provided.

Referee Report

4 major / 5 minor

Summary. The paper studies how pretraining imbalance in foundation models affects long-tailed downstream classification. It distinguishes parameter imbalance (bias inherited from imbalanced pretraining data, estimated through a GLA-based prior) from data imbalance (bias from the downstream long-tailed training set), and argues empirically that parameter imbalance is more influential and cannot be fixed by logit adjustment. The proposed method, framed as a backdoor adjustment, treats an 'incomplete semantic factor' C as a confounder and averages LA-adjusted predictions from CLIP, OpenCLIP, and MetaCLIP. Experiments on ImageNet-LT, Places365-LT, and iNaturalist2018 show consistent accuracy gains over PEFT baselines, with additional ablations over M and PEFT variants.

Significance. The paper contains a useful and reproducible empirical study: the finding that logit-adjustment improves the classifier more than the representation (Tab. 4), and the observation that GLA-Train does not remove the pretraining bias (Tab. 3), are informative for practitioners. The proposed method is simple, code is released, and the reported gains are consistent across three large long-tailed benchmarks and two PEFT methods. However, the central advertised contribution is the causal claim that the method 'learns the true causal effect between input samples and labels.' That claim is not supported: the estimator in Eq. (7) reduces to an equal-weight ensemble of LA-adjusted models, and no identification argument for C or the uniform prior is provided. The empirical method may still be valuable as a diversified ensemble, but the causal interpretation is the load-bearing part of the paper's framing and is currently unsubstantiated.

major comments (4)

[Sec. 4.2, Eq. (7)] The backdoor adjustment in Eq. (7) requires C to be a well-defined confounder with a known distribution, but no constructive definition or measurement model for C is given. The three foundation models are treated as instantiations of C, yet model identity is not a value of a semantic factor; the Grad-CAM evidence in Fig. 4 and Sec. E.6 shows differential attention, which demonstrates diversity, not identification. Consequently, with P(c)=1/M, Eq. (7) is an equal-weight average of LA-adjusted logits, and the Abstract's claim that the method 'learns the true causal effect' is not established. A necessary control is a baseline that applies the same LA adjustment to each of the three models and simply averages their logits, without the causal framing.
[Sec. 4.2 and Tab. 8] The uniform prior P(c)=1/M and the exchangeability of CLIP, OpenCLIP, and MetaCLIP are load-bearing assumptions, but they are not tested or defended. The M-ablation in Tab. 8 only shows that adding more foundation models improves accuracy, which is also the generic behavior of an ensemble and is compatible with a model in which C has no causal role. Without a check of the prior or an identification argument for C, the method cannot be described as a backdoor adjustment; it is an ensemble with equal weights.
[Sec. 3.1, Def. 3.1, Eq. (3)] The empirical decomposition into parameter imbalance and data imbalance relies entirely on the GLA-estimated prior bPP(Y), both for the P-Many/P-Medium/P-Few groupings and for Definition 3.1. Since bPP(Y) is fitted on the validation set (Eq. 3) and the true PP(Y) is inaccessible, the paper should provide a sensitivity analysis showing that the qualitative conclusions in Fig. 3 and Tab. 2 are robust to errors in bPP(Y); otherwise the central claim that parameter imbalance 'plays a more critical role' could be an artifact of the estimator rather than a property of pretraining data.
[Appendix E.9] The synthetic experiment intended to verify that C is a confounder defines C explicitly as an object part in generated images, so it cannot validate the real-world identification of C across CLIP, OpenCLIP, and MetaCLIP. Moreover, the statement that a model trained on Atrain estimates P(Y|do(X)) is asserted rather than derived; balanced training over C is not an intervention in the causal sense, so the experiment does not supply the missing identification argument.

minor comments (5)

[Abstract and Sec. 5] The claimed 'average performance increase of about 1.67%' is not precisely defined; please specify which baselines and datasets are averaged, and report variance or significance if available.
[Appendix A.1] The line 'It contains a total of 12.21K images' for ImageNet-LT is a factual error; ImageNet-LT has approximately 115.8K training images.
[Tab. 5] The Decoder row reports '21.26 34' under Params and Epochs; this appears to be a formatting error with missing units (presumably 21.26M and 34).
[Appendix E.7] The sentence 'As shown in Fig. 8, the result verifies our point' appears to refer to Fig. 9, which is the figure showing the comparison between training from scratch and fine-tuning.
[Tabs. 3, 11, 13] There are several typographical errors in table captions and text, including 'Ther performance' in Tabs. 3 and 11 and 'Resverse' in Tab. 13; these should be corrected.

Circularity Check

1 steps flagged · score 6.0 of 10

The backdoor adjustment in Eq. (7) is an equal-weight ensemble over three open checkpoints; the 'true causal effect' is defined, not derived, so the causal claim is circular even though the benchmark gains are empirical.

renaming known result [Sec. 4.2, Eq. (7); Sec. 4.2 'Backdoor adjustment'; Ablation Tab. 8]
"P(Y=y|do(x)) = Σ_{b∈B} Σ_{c∈C} P(Y=y|b,c)P(c)P(b|x) = Σ_{c∈C} P(Y=y|b,c)P(c) (7) ... For simplicity, we assume P(c) = 1/M, where M is the number of incomplete semantic factors. ... Since iterating over all possible incomplete semantic factors is impractical, we approximate these factors using models like CLIP, OpenCLIP, and MetaCLIP, thereby addressing Eq. 7 and simplifying the process of balancing both parameter and data imbalances simultaneously."

Eq. (7) is the complete estimator: the injective B mapping gives P(b|x)=1, so P(Y|do(x)) is defined as Σ_c P(Y|b,c)P(c), and with P(c)=1/M the only operational values of c are the three checkpoints CLIP, OpenCLIP, and MetaCLIP. Hence the 'backdoor-adjusted' score is by construction the equal-weight average of the LA-debiased outputs of those three models. The paper supplies no independent measurement of C, no argument that the three checkpoints are draws from a distribution over incomplete semantic factors, and no test of the uniform prior; the causal interpretation is attached to an ensemble average rather than derived from identification. The monotone M-ablation (Tab.

full rationale

The paper's empirical study of parameter vs data imbalance (Sec. 3) is not circular: it measures accuracy gaps across checkpoints, uses an external GLA prior estimate, and contrasts CE/LA/GLA-Train; those findings can stand on their own. Similarly, the reported 1.67% average gains are genuine benchmark numbers, and the method is self-contained against external baselines. The circularity is confined to the causal claim. Section 4.2 labels a uniform average of LA-adjusted logits from CLIP, OpenCLIP, and MetaCLIP as a backdoor adjustment, assumes P(c)=1/M without identification, and then uses the M-ablation as confirmation. That is a renaming of an equal-weight ensemble into causal language. Since the central advertised contribution is 'learns the true causal effect,' this is partial circularity (score 6): the empirical component is independent, but the causal interpretation is definitional. No load-bearing self-citation chain is present; no fitted parameter is renamed as a prediction beyond the causal label applied to Eq. (7).

Assumptions & free parameters 2 free parameters · 5 assumptions · 3 invented entities

The central method rests on a latent confounder C that is approximated by three pretrained models, a uniform prior over C, and an injective X to B mapping. None of these are derived; they are modeling choices. The GLA-estimated prior is fitted to the validation set and is load-bearing for the parameter-imbalance analysis. The paper's empirical gains are real correlations, but the causal interpretation is not identified.

free parameters (2)

M (number of incomplete semantic factors) = 3 (CLIP, OpenCLIP, MetaCLIP)
M is chosen because three CLIP variants are available; Table 8 shows performance rises with M, so it is an ad hoc hyperparameter, not derived from the causal model.
Estimated pre-training label prior from GLA (Eq. 3) = per-class estimated priors
Used to define parameter imbalance groups (P-Many/Medium/Few) and to motivate the method. It is fitted by constrained optimization on the validation set and is never checked against the true but inaccessible pre-training distribution.

assumptions (5)

domain assumption Incomplete semantic factor C is a confounder with fork X <- C -> Y (Fig. 5).
The causal graph is asserted in Sec. 4.1; Appendix E.9 provides a synthetic Stable Diffusion experiment as support, but C is not measured in real data.
domain assumption The mapping X -> B is injective and P(b|x) = 1 for the unique balanced representation b (Eq. 7).
Assumed in Sec. 4.2 to drop the sum over B; no proof or definition of the balanced representation space is given.
ad hoc to paper P(c) = 1/M, uniform over incomplete semantic factors.
Stated as a simplification in Sec. 4.2; no estimation or justification from data.
ad hoc to paper CLIP, OpenCLIP, and MetaCLIP are exchangeable instantiations of C.
The method approximates the sum over C by using three fixed pretrained models; there is no argument that these three span or sample the confounder space.
domain assumption GLA's estimated prior is a valid proxy for the true pre-training label prior.
Definitions 3.1 and the P-Many/Medium/Few groupings rely on the GLA estimate from Eq. (3), borrowed from Zhu et al. 2024; validity is assumed.

invented entities (3)

Incomplete semantic factor C
purpose: Latent confounder that creates spurious correlation between input X and label Y, justifying backdoor adjustment.
C is never measured. It is identified with whatever semantic parts the three chosen models attend to; no falsifiable handle outside the paper.
Parameter imbalance D (causal variable)
purpose: Postulated latent cause in the graph D -> C that represents pre-training label bias in the model parameters.
D is inferred from the GLA-estimated prior and is not independently observable.
Inaccessible semantic factor U
purpose: Explains data generation X <- U and makes the semantic factor incomplete.
Introduced in Fig. 5 and Sec. B as a placeholder; not used in the method.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Rethinking the Bias of Foundation Model under Long-tailed Distribution." pith.science (2026). https://pith.science/paper/AFUEXSQA

@misc{pith2026250115955,
  author       = {Pith},
  title        = {Pith review of: Rethinking the Bias of Foundation Model under Long-tailed Distribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AFUEXSQA}},
  note         = {Machine review of arXiv:2501.15955}
}

abstract

Long-tailed learning has garnered increasing attention due to its practical significance. Among the various approaches, the fine-tuning paradigm has gained considerable interest with the advent of foundation models. However, most existing methods primarily focus on leveraging knowledge from these models, overlooking the inherent biases introduced by the imbalanced training data they rely on. In this paper, we examine how such imbalances from pre-training affect long-tailed downstream tasks. Specifically, we find the imbalance biases inherited in foundation models on downstream task as parameter imbalance and data imbalance. During fine-tuning, we observe that parameter imbalance plays a more critical role, while data imbalance can be mitigated using existing re-balancing strategies. Moreover, we find that parameter imbalance cannot be effectively addressed by current re-balancing techniques, such as adjusting the logits, during training, unlike data imbalance. To tackle both imbalances simultaneously, we build our method on causal learning and view the incomplete semantic factor as the confounder, which brings spurious correlations between input samples and labels. To resolve the negative effects of this, we propose a novel backdoor adjustment method that learns the true causal effect between input samples and labels, rather than merely fitting the correlations in the data. Notably, we achieve an average performance increase of about $1.67\%$ on each dataset. Code is available: https://github.com/JiahaoChen1/Pre-train-Imbalance

Figures

Figures reproduced from arXiv: 2501.15955 by the authors.

**Figure 1.** Previous methods focus on using the downstream data to fine-tune the foundation model while ignoring that the pre-training data has a potential influence of bias (dashed line). directly affects the downstream task, which we define as data imbalance. Through fine-tuning, we find that both types of imbalance influence downstream tasks, but parameter imbalance plays a more significant role, as shown in [PITH_FULL_IMA… view at source ↗

**Figure 2.** The performance of different groups with (a) CE and (b) LA on Places365-LT dataset. training set by introducing a scaling factor. PT (Y = y | x) = PT (x | Y = y)PT (Y = y) PT (x) ∝ PS(x | Y = y)PT (Y = y) ∝ PT (Y = y) PS(Y = y) PS(Y = y | x) (1) There are two types of LA, either applied post-hoc to a trained model or enforced in the loss during training, and the latter can achieve better performance. When integratin… view at source ↗

**Figure 3.** The data and parameter imbalance on (a) ImageNet-LT and (b) Places365-LT. The class indices of the left picture are sorted relying on the data imbalance while the right picture relies on the parameter imbalance. Curves are smoothed for better visualization. Parameter imbalance occupies a more vital role [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (8 more)

**Figure 5.** Figure 5: The framework of our proposed method. across various classes. 4.1. A causal view for long-tailed learning We model the long-tailed image classification process with a causal structure graph as shown in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: The performance of different groups with our method on Places365-LT. Backdoor adjustment relieves the parameter imbalance. To verify that our method effectively alleviates parameter imbalance, we report the performance results in [PITH_FULL_IMAGE:figures/full_fig_p…

**Figure 7.** Figure 7: We constitute different parameter imbalanced datasets (a) the original Places365-LT, (b) the data imbalance and parameter imbalance are consistent, (c) the data imbalance and parameter imbalance are reversed, and (d) the downstream dataset is balanced. Each dot in the …

**Figure 8.** Figure 8: Prediction scores of fine-tuned OpenCLIP and MetaCLIP. OpenCLIP is more confident about the sample if the head is exposed (C = 0), while MetaCLIP is more sensitive to the body (C = 1). the Consistency scenario. In the Reverse setting, the tail classes (in terms of para…

**Figure 9.** Figure 9: The comparison between fine-tuned model and training from scratch. To show the existence of confounding bias directly, we measure differences between P(Y | X) and P(Y | do(X)). As shown in Tab. 16, there is a significant difference between P(Y | X) and P(Y | do(X), whi…

**Figure 10.** Figure 10: The comparison of our method with other fine-tuning based methods [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]

**Figure 11.** Figure 11: The visualization of the “dog” class: C = 0 represents samples generated for the dog’s head, while C = 1 corresponds to samples of the dog’s body. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]

**Figure 12.** Figure 12: The performance curve illustrates four evaluation scenarios: Atrain, Atest indicates the model is trained on Atrain and evaluated on Atest; Atrain, Btest represents the model trained on Atrain and evaluated on Btest; Btrain, Atest corresponds to the model trained on B…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 12 canonical work pages

[1]

Adaptformer: Adapting vision transformers for scalable visual recognition

Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., and Luo, P. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35: 0 16664--16678, 2022

2022
[2]

Reproducible scaling laws for contrastive language-image learning

Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., and Jitsev, J. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 2818--2829, 2023

2023
[3]

and Hart, P

Cover, T. and Hart, P. Nearest neighbor pattern classification. IEEE transactions on information theory, 13 0 (1): 0 21--27, 1967

1967
[4]

Class-balanced loss based on effective number of samples

Cui, Y., Jia, M., Lin, T.-Y., Song, Y., and Belongie, S. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 9268--9277, 2019

2019
[5]

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. Ieee, 2009

2009
[6]

Lpt: Long-tailed prompt tuning for image classification

Dong, B., Zhou, P., Yan, S., and Zuo, W. Lpt: Long-tailed prompt tuning for image classification. arXiv preprint arXiv:2210.01033, 2022

arXiv 2022
[7]

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

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

arXiv 2010
[8]

and Wang, S

Guo, H. and Wang, S. Long-tailed multi-label visual recognition by collaborative training on uniform and re-balanced samplings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 15089--15098, 2021

work page 2021

Show all 40 references

[9]

Deep residual learning for image recognition

He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 770--778, 2016

2016
[10]

Parameter-efficient transfer learning for nlp

Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., and Gelly, S. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pp.\ 2790--2799. PMLR, 2019

2019
[11]

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

Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021

2021 arXiv
[12]

Visual prompt tuning

Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., and Lim, S.-N. Visual prompt tuning. In European Conference on Computer Vision, pp.\ 709--727. Springer, 2022

2022
[13]

Decoupling representation and classifier for long-tailed recognition

Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., and Kalantidis, Y. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019

1910 arXiv
[14]

D., Jeong, J., and Kim, G

Kim, C. D., Jeong, J., and Kim, G. Imbalanced continual learning with partitioning reservoir sampling. In European Conference on Computer Vision, pp.\ 411--428. Springer, 2020

2020
[15]

Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., and Yu, S. X. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 2537--2546, 2019

2019
[16]

Retrieval augmented classification for long-tail visual recognition

Long, A., Yin, W., Ajanthan, T., Nguyen, V., Purkait, P., Garg, R., Blair, A., Shen, C., and van den Hengel, A. Retrieval augmented classification for long-tail visual recognition. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 6949--6959, ...

2022
[17]

A simple long-tailed recognition baseline via vision-language model

Ma, T., Geng, S., Wang, M., Shao, J., Lu, J., Li, H., Gao, P., and Qiao, Y. A simple long-tailed recognition baseline via vision-language model. arXiv preprint arXiv:2111.14745, 2021

2021 arXiv
[18]

K., Jayasumana, S., Rawat, A

Menon, A. K., Jayasumana, S., Rawat, A. S., Jain, H., Veit, A., and Kumar, S. Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314, 2020

2007 arXiv
[19]

Causality

Pearl, J. Causality. Cambridge university press, 2009

2009
[20]

W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp.\ 8748--8763. PMLR, 2021

2021
[21]

Balanced meta-softmax for long-tailed visual recognition

Ren, J., Yu, C., Ma, X., Zhao, H., Yi, S., et al. Balanced meta-softmax for long-tailed visual recognition. Advances in Neural Information Processing Systems, 33: 0 4175--4186, 2020

2020
[22]

High-resolution image synthesis with latent diffusion models

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

2022
[23]

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

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. arXiv preprint arXiv:2111.02114, 2021

2021 arXiv
[24]

R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., and Batra, D

Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., and Batra, D. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pp.\ 618--626, 2017

2017
[25]

Parameter-efficient long-tailed recognition

Shi, J.-X., Wei, T., Zhou, Z., Han, X.-Y., Shao, J.-J., and Li, Y.-F. Parameter-efficient long-tailed recognition. arXiv preprint arXiv:2309.10019, 2023

2023 arXiv
[26]

Long-tail learning with foundation model: Heavy fine-tuning hurts

Shi, J.-X., Wei, T., Zhou, Z., Shao, J.-J., Han, X.-Y., and Li, Y.-F. Long-tail learning with foundation model: Heavy fine-tuning hurts. In Forty-first International Conference on Machine Learning, 2024

2024
[27]

and Seo, S.-W

Suh, M.-K. and Seo, S.-W. Long-tailed recognition by mutual information maximization between latent features and ground-truth labels. In International Conference on Machine Learning, pp.\ 32770--32782. PMLR, 2023

2023
[28]

Long-tailed classification by keeping the good and removing the bad momentum causal effect

Tang, K., Huang, J., and Zhang, H. Long-tailed classification by keeping the good and removing the bad momentum causal effect. Advances in neural information processing systems, 33: 0 1513--1524, 2020

2020
[29]

Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition

Tian, C., Wang, W., Zhu, X., Dai, J., and Qiao, Y. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. In European Conference on Computer Vision, pp.\ 73--91. Springer, 2022

2022
[30]

The inaturalist species classification and detection dataset

Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., and Belongie, S. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 8769--8778, 2018

2018
[31]

Exploring vision-language models for imbalanced learning

Wang, Y., Yu, Z., Wang, J., Heng, Q., Chen, H., Ye, W., Xie, R., Xie, X., and Zhang, S. Exploring vision-language models for imbalanced learning. International Journal of Computer Vision, 132 0 (1): 0 224--237, 2024

2024
[32]

Dualprompt: Complementary prompting for rehearsal-free continual learning

Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.-Y., Ren, X., Su, G., Perot, V., Dy, J., et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pp.\ 631--648. Springer, 2022 a

2022
[33]

Learning to prompt for continual learning

Wang, Z., Zhang, Z., Lee, C.-Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., and Pfister, T. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 139--149, 2022 b

2022
[34]

Generalization beyond data imbalance: A controlled study on clip for transferable insights

Wen, X., Zhao, B., Chen, Y., Pang, J., and Qi, X. Generalization beyond data imbalance: A controlled study on clip for transferable insights. arXiv preprint arXiv:2405.21070, 2024

2024 arXiv
[35]

E., Huang, P.-Y., Howes, R., Sharma, V., Li, S.-W., Ghosh, G., Zettlemoyer, L., and Feichtenhofer, C

Xu, H., Xie, S., Tan, X. E., Huang, P.-Y., Howes, R., Sharma, V., Li, S.-W., Ghosh, G., Zettlemoyer, L., and Feichtenhofer, C. Demystifying clip data. arXiv preprint arXiv:2309.16671, 2023 a

2023 arXiv
[36]

Learning imbalanced data with vision transformers

Xu, Z., Liu, R., Yang, S., Chai, Z., and Yuan, C. Learning imbalanced data with vision transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023 b

2023
[37]

B., Ravfogel, S., and Goldberg, Y

Zaken, E. B., Ravfogel, S., and Goldberg, Y. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199, 2021

2021
[38]

Generalized logit adjustment: Calibrating fine-tuned models by removing label bias in foundation models

Zhu, B., Tang, K., Sun, Q., and Zhang, H. Generalized logit adjustment: Calibrating fine-tuned models by removing label bias in foundation models. Advances in Neural Information Processing Systems, 36, 2024

2024
[39]

P., and Jiang, Y.-G

Zhu, J., Wang, Z., Chen, J., Chen, Y.-P. P., and Jiang, Y.-G. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 6908--6917, 2022

2022
[40]

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 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Adaptformer: Adapting vision transformers for scalable visual recognition

Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., and Luo, P. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35: 0 16664--16678, 2022

2022

[2] [2]

Reproducible scaling laws for contrastive language-image learning

Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., and Jitsev, J. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 2818--2829, 2023

2023

[3] [3]

and Hart, P

Cover, T. and Hart, P. Nearest neighbor pattern classification. IEEE transactions on information theory, 13 0 (1): 0 21--27, 1967

1967

[4] [4]

Class-balanced loss based on effective number of samples

Cui, Y., Jia, M., Lin, T.-Y., Song, Y., and Belongie, S. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 9268--9277, 2019

2019

[5] [5]

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. Ieee, 2009

2009

[6] [6]

Lpt: Long-tailed prompt tuning for image classification

Dong, B., Zhou, P., Yan, S., and Zuo, W. Lpt: Long-tailed prompt tuning for image classification. arXiv preprint arXiv:2210.01033, 2022

arXiv 2022

[7] [7]

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

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

arXiv 2010

[8] [8]

and Wang, S

Guo, H. and Wang, S. Long-tailed multi-label visual recognition by collaborative training on uniform and re-balanced samplings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 15089--15098, 2021

work page 2021