Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Multimodal Approaches to Fair Image Classification: An Ethical Perspective

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A weighted blend of generated-image and text embeddings lifts zero-shot classification accuracy and can offset demographic bias.

desk verdict Honest, clearly written thesis with a fixable but serious evaluation flaw: test-set-selected weights drive the headline gains, so the numbers need re-estimation before the fairness claims can be credited. read the letter →

arxiv 2412.12165 v1 pith:ISAQV6J7 submitted 2024-12-11 cs.CV cs.AIcs.CYcs.LG

Javon Hickmon This is my paper · ORCID
classification cs.CVcs.AIcs.CYcs.LG
keywords multimodalimageclassificationzero-shotCLIPembeddingsStableDiffusionsyntheticdatagenerationdemographicbiasfairnessinference-timeensembling
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 thesis tries to establish that a training-free, inference-time fix can make pretrained multimodal classifiers both more accurate and less demographically biased. The fix is to generate, for each class, a descriptive prompt and several synthetic images, embed both in the model's shared space, and replace the pure text embedding with a weighted sum of the text and image embeddings before measuring cosine similarity to the query image. If the generated images represent the true class distribution better than the class name does, and if the base model can tell similar classes apart, the combined embedding lands closer to the query image and classification improves. The author reports accuracy gains of roughly 1-3 points across four diverse benchmarks, a correction of fine-grained confusions, and a 4-7 point gain on race classification, with the explicit caveat that the method offsets learned biases rather than removing them.

What carries the argument

The load-bearing object is the weighted combined embedding $f(w) = w\mathbf{t} + (1-w)\mathbf{i}$, where $\mathbf{t}$ is the normalized text embedding for a class (a CuPL-generated description or a template prompt) and $\mathbf{i}$ is the average of the normalized embeddings of several synthetic images of that class, generated from the same text by Stable Diffusion XL. The query image's normalized embedding $\mathbf{q}$ is classified by cosine similarity $f(w)\cdot\mathbf{q}$, and the mixture weight $w\in[0,1]$ is either scanned in 0.01 steps to maximize top-1 accuracy or fixed at a small value such as $w=0.1$. The same mechanism powers MuSE and D3G; D3G only changes the prompts to include demographic attributes so that generated images sample the class conditionally on race, gender, or age.

What would settle it

On any benchmark where the generative model's synthetic images clearly mismatch the test distribution, fix the weight $w$ using a validation set before seeing test labels, then compare against text-only CLIP; if accuracy drops on those mismatched classes, the claim that the combined embedding lands closer to the true query embedding is falsified. The RESISC45 airplane case, where generated images show planes in flight but ground truth shows parked planes, is a natural place to run this test.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims two things. First, MuSE (Multimodal Synthetic Embeddings) improves the accuracy and generalizability of zero-shot multimodal classifiers at inference time by taking, for each class, a weighted sum of a text embedding and the averaged embedding of several Stable Diffusion XL images generated from CuPL descriptions; classifying by cosine similarity against this combined embedding beats text-only CLIP on Flowers 102, DTD, FGVC Aircraft, and RESISC45, and it corrects cases where CLIP collapses fine-grained classes together. Second, D3G (Diverse Demographic Data Generation) applies the same recipe with prompts that vary race, gender, and age, and reports that the demographic-diverse synthetic images boost profession and race classification on the IdenProf dataset while shifting per-class accuracies toward underrepresented groups. The paper is explicit that the method does not remove bias; it offsets learned bias, so it can also amplify bias if the generated content is biased.

Load-bearing premise

The method works only if the generated class images sit closer to the real test images than the text prompt alone does, and the base model can already tell similar classes apart; if either fails, the weighted sum pushes predictions in a direction that is not trustworthy.

Editorial extensions

If this is right

  • If the weighted-sum mechanism is correct, any pretrained multimodal model can be upgraded at inference time with no training or fine-tuning, as long as a text-to-image model and a prompt generator are available.
  • Fine-grained classes that the base model confuses should recover: the mushroom/agaric and frilled lizard/agama examples show the combined embedding rebalances per-class accuracy instead of overfitting one class.
  • Demographic classification in zero-shot settings can be improved by 4-7 points simply by generating class images across demographic groups, without retraining the classifier.
  • Because the method only offsets bias, deployment choices matter: using it in contexts where generated images carry stereotypes can worsen unfair outcomes, as the ethics statement warns.
  • Classes the base model has no cross-modal representation of (the Bishop of Llandaff example, 0% baseline) can be rescued when the prompt is enriched with compositional detail, suggesting prompt quality is part of the mechanism.

Reading between the lines

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

  • A direct test the paper leaves implicit: choose the weight $w$ on a validation split rather than by scanning to maximize accuracy on the test set; if the reported gains shrink or vanish, part of the improvement is an artifact of test-set weight selection.
  • The same recipe should transfer to other shared-embedding models and modalities: any open-vocabulary model with separate encoders plus any generative model could use the identical weighted-sum construction, so the claim is testable beyond CLIP and SDXL.
  • The paper's per-class D3G result that adding a majority demographic (White) can lower White classification accuracy suggests the embedding shift is not monotone; a natural extension is to weight generated images inversely to their representation in the training distribution, an idea the author flags for future work.
  • If the mechanism is really about landing closer to the true query embedding, then the method's benefit should correlate with how well the generative model's samples match the test distribution; this predicts that on distribution-shifted benchmarks the gains will concentrate only on classes whose synthetic images are faithful.
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.

Referee Report

3 major / 6 minor

Summary. The thesis proposes two training-free, zero-shot inference-time methods. MuSE combines CuPL-generated text prompts with SDXL-generated images: for each class the normalized embeddings of five generated images are averaged and then combined with the text embedding via a weighted sum, with the weight w scanned from 0 to 1. D3G applies the same idea to demographic attributes (race, gender, age) and evaluates on IdenProf images labeled with FairFace-derived demographics. The paper reports 1-3% accuracy gains for MuSE and 4-7% gains for D3G on Race 7 classification, plus a qualitative fine-grained analysis and an ethics discussion.

Significance. The central idea of using a generative model to construct an inference-time support set for a frozen vision-language classifier is simple and potentially useful, and the thesis is commendably explicit about the assumptions behind the method (Section 2.6 and Section 3.6) and about its limitations (Section 3.9). The ethics statement (Section 3.8) also engages seriously with misuse and stereotype reinforcement. However, the empirical evidence as presented does not yet support the headline claims: the accuracy gains are obtained by selecting the interpolation weight on the test set used for reporting, and the D3G fairness claim is not measured with any standard bias metric. With a held-out validation protocol and proper bias metrics, the method could be a useful contribution to training-free multimodal classification.

major comments (3)
  1. [Sections 2.5.2 and 3.4.4] The main accuracy result is fitted, not predicted. Section 2.5.2 states that after scanning weights from 0 to 1 in steps of 0.01, "we take the maximum top-1 accuracy and use that weight configuration for our final result"; the same procedure is described in Section 3.4.4, and Table 3.4 reports the resulting weights. Because the test set is used both to select w and to report accuracy, the reported MuSE and D3G gains (including the 4-7% D3G Race 7 gains) are optimistic upper bounds. The paper needs held-out weight selection (e.g., a validation split or nested cross-validation) or a pre-specified weight with a sensitivity analysis; the current text provides neither.
  2. [Abstract, Sections 3.1 and 3.7] The claim that D3G "reduces demographic bias" is unsupported because no bias metric is defined or reported. Section 3.5.1 uses only top-1 accuracy, and Section 3.5.5 reports per-class accuracies. Per-class accuracy on the demographic classification task is not a measure of demographic bias for the actual use case (e.g., profession classification). For example, under the "Race 7" prompt in Table 3.5, per-class accuracy is 8.92% for White and 66.42% for Black; these numbers describe absolute performance, not fairness. To support the fairness claim, the authors should report demographic parity, equalized odds, accuracy gaps on a downstream task with protected attributes, or another standard fairness metric.
  3. [Sections 2.5.2 and 2.5.3, Table 2.5] The "w = 0.1" rows are presented as a fixed-weight robustness check, but the text itself shows that 0.1 was not specified before seeing test results. Section 2.5.3 notes that for two evaluations "MuSE Standard where w = 0.1 ... ended up being the optimal configuration," i.e., the fixed value was identified from the same test-set scan. Moreover, Table 2.5 shows the fixed weight can hurt performance (RESISC45: 71.06 vs. baseline 71.45), so the method as currently specified does not provide a principled way to choose w without test labels.
minor comments (6)
  1. [Section 2.5.2] The mathematical definitions of xStd and xConf are not well-formed; for example, "f(w) = {f(w) | ...}" is self-referential and should be rewritten as a clear argmax over w.
  2. [Table 2.1] Table 2.1 mixes Mean Per-Class Accuracy and Top-1 Accuracy; cross-dataset comparisons should state explicitly that these metrics are not directly comparable.
  3. [Section 2.5.4] The fine-grained analysis is hard to interpret without the number of test images and the exact class-selection criterion; for example, Table 2.6 reports "0.1" and "1.0" as accuracies, which should be clarified as fractions over how many examples.
  4. [Throughout] There are numerous typos and grammatical errors: acknowledgements "I would like to my mentor," Section 2.2 "in the field in the field," Section 3.5.5 "demograpics" and "this my imply," and Appendix A.2 "Cardoworking." These should be corrected in a revision.
  5. [Section 3.2 and Figure 3.2] The Google Photos "gorillas" incident is described without a citation; please add a source for this widely reported but still citable claim.
  6. [Appendix B] Appendix B contains duplicated table captions and malformed prompts (e.g., "A photo of a<age> year old"), which should be corrected for reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported MuSE and D3G accuracy gains come from test-set-selected interpolation weights, making the headline numbers fitted optima rather than out-of-sample predictions.

  1. fitted input called prediction [§2.4.3 and §2.5.2 (MuSE architecture and evaluation breakdown)]
    "We determine the weight for this sum, by scanning values from 0 to 1 by increments of 0.01, where the text is multiplied by weight w and the images are multiplied by weight (1 − w). ... Upon scanning all values, we take the maximum top-1 accuracy and use that weight configuration for our final result."

    The reported 'MuSE Standard' accuracy is obtained by choosing w as the argmax of top-1 accuracy computed on the same test set used for the results tables. The baseline CLIP is not given this selection step. Therefore the reported improvement is, by construction, the maximum over the scanned weight family rather than the accuracy of a fixed, prespecified method. The fixed w=0.1 rows avoid the in-table argmax, but §2.5.3 states that this value was inferred from the evaluations rather than pre-registered, so the central claim still reduces to test-set selection.

  2. fitted input called prediction [§3.4.4 and §3.5.3–3.5.4 (D3G weighted sum and results)]
    "Upon getting these embeddings, we scan values from 0 to 1 using a step value of 0.01 in order to find an optimal weight to create a weighted sum of the text and image embeddings."

    D3G reports its accuracy gains, including the 4-7% improvement on Race 7, using weights that were optimized on the same test set whose top-1 accuracy is reported. Table 3.4 lists the selected weights after evaluation. The headline D3G accuracy improvement is therefore a fitted maximum over w rather than a prediction at a fixed hyperparameter, so the same reduction applies as for MuSE Standard.

full rationale

Both central accuracy claims reduce by construction to test-set weight selection. In §2.5.2, MuSE Standard reports the argmax over w of top-1 accuracy computed on the evaluation set; in §3.4.4, D3G does the same. The baseline CLIP has no such tuning, so the reported 1-7% gains are upper bounds of the scanned family rather than predictions at a fixed weight. This is the fitted-input-called-prediction pattern. The fixed w=0.1 rows in Tables 2.2-2.5 provide partial independent evidence, but §2.5.3 explicitly says the value was chosen based on the specific evaluation, so it is not a prespecified hyperparameter. The per-class D3G tables are informative but cannot support the headline claim that demographic bias was reduced, because no standard bias metric such as demographic parity, equalized odds, or accuracy gap is reported; this is a measurement gap rather than circularity. I found no load-bearing self-citation: CuPL is cited as a published external method and is used as a prompt generator, not to justify the evaluation claims, and no uniqueness argument is imported. Overall score 6: the strongest reported numbers are selected optima rather than predictions, so the central claim is partially circular, while the method itself has independent content that could be validated with held-out weights.

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

The central method rests on fitted interpolation weights and on assumptions about generative and multimodal models that the authors explicitly list in Sec 2.6 and Sec 3.6. The fitted weight w is the main free parameter, and several other hyperparameters are hand-chosen without ablations. No new physical or mathematical entities are introduced; MuSE and D3G are algorithmic procedures, not invented entities.

free parameters (4)
  • w (text-image interpolation weight) = Varies per dataset and evaluation, e.g., 0.85/0.15 for D3G profession, 0.68/0.32 for D3G race 7 (Table 3.4)
    Chosen by scanning 0 to 1 in steps of 0.01 and selecting the value that maximizes accuracy on the evaluation set (Sec 2.5.2 and Sec 3.4.4). This is a test-set-fitted parameter.
  • Number of generated images averaged per class = 5 for MuSE; 1 (standard) or 5 (average image) for D3G
    Hand-chosen in Sec 2.4.3 and Sec 3.4.3; no ablation is provided for this choice, and it directly affects the averaged image embedding.
  • SDXL generation settings = 50 steps, guidance scale 15, seed 0
    Hand-set in Sec 2.4.3; no sensitivity analysis is reported.
  • Number of CuPL prompts averaged = 1 or 20 to 60 depending on the dataset and evaluation
    Chosen per evaluation in Sec 2.5.3; it changes the text embedding and affects whether MuSE improves or fails on RESISC45.
assumptions (4)
  • domain assumption The generative model has a better learned representation of the true data distribution than the base text embedding alone.
    Stated in Sec 2.6 and Sec 3.6 as a required assumption for MuSE and D3G to improve accuracy.
  • domain assumption The base multimodal model can distinguish similar classes in embedding space; otherwise the weighted sum does not move toward the true embedding.
    Stated in Sec 2.6 and Sec 3.6; the authors explicitly say the method will not improve performance if classes map to the same embedding.
  • domain assumption SDXL-generated images from class descriptions are distributed similarly enough to real query images for their averaged CLIP embeddings to be useful.
    Implicit in the MuSE architecture (Sec 2.4.3); the RESISC45 airplane failure described in Sec 2.5.3 is an admitted counterexample.
  • domain assumption FairFace classifier labels for IdenProf demographics are accurate enough to serve as ground truth for evaluating D3G.
    Demographic labels are assigned by a pretrained FairFace classifier (Sec 3.4.1); the authors note annotator bias in the FairFace dataset in Sec 3.8.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Approaches to Fair Image Classification: An Ethical Perspective." pith.science (2026). https://pith.science/paper/ISAQV6J7

@misc{pith2026241212165,
  author       = {Pith},
  title        = {Pith review of: Multimodal Approaches to Fair Image Classification: An Ethical Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ISAQV6J7}},
  note         = {Machine review of arXiv:2412.12165}
}
read the original abstract

In the rapidly advancing field of artificial intelligence, machine perception is becoming paramount to achieving increased performance. Image classification systems are becoming increasingly integral to various applications, ranging from medical diagnostics to image generation; however, these systems often exhibit harmful biases that can lead to unfair and discriminatory outcomes. Machine Learning systems that depend on a single data modality, i.e. only images or only text, can exaggerate hidden biases present in the training data, if the data is not carefully balanced and filtered. Even so, these models can still harm underrepresented populations when used in improper contexts, such as when government agencies reinforce racial bias using predictive policing. This thesis explores the intersection of technology and ethics in the development of fair image classification models. Specifically, I focus on improving fairness and methods of using multiple modalities to combat harmful demographic bias. Integrating multimodal approaches, which combine visual data with additional modalities such as text and metadata, allows this work to enhance the fairness and accuracy of image classification systems. The study critically examines existing biases in image datasets and classification algorithms, proposes innovative methods for mitigating these biases, and evaluates the ethical implications of deploying such systems in real-world scenarios. Through comprehensive experimentation and analysis, the thesis demonstrates how multimodal techniques can contribute to more equitable and ethical AI solutions, ultimately advocating for responsible AI practices that prioritize fairness.

Figures

Figures reproduced from arXiv: 2412.12165 by the authors.

Figure 2.1
Figure 2.1. The CuPL Architecture [Pratt et al., 2023] Multimodal Prompting with Missing Modalities for Visual Recognition [Lee et al., 2023] attempts to reduce the effect of missing modalities on classification accuracy, an issue that can introduce bias into multimodal transformers. Despite this, a significant limitation was that they could not recover missing information from the multimodal inputs. We intend to solve this iss… view at source ↗
Figure 2.2
Figure 2.2. Framework to Handle Missing Modalities [Lee et al., 2023] A model that applies this concept of model ensembling was introduced in Learning to Navigate for Fine￾grained Classification by Yang et al. [2018]. This is a state-of-the-art paper that attempts to reduce misclas￾sification rates by developing a model called NTS-Net (Navigator-Teacher-Scrutinizer Network) to teach itself methods of identifying and scrutinizin… view at source ↗
Figure 2.3
Figure 2.3. The NTSNet Architecture [Yang et al., 2018] 2.4 Methods In this section, we explore and analyze our selection of datasets, models, the MuSE Architecture, and implications of the final results. 2.4.1 Datasets Within this chapter, we specifically select four datasets that will test the generalizability and fine-grained classification capabilities of MuSE: Flowers 102 [Nilsback and Zisserman, 2008], Describable Texture… view at source ↗
Figures from the paper (11 more)
Figure 2.4
Figure 2.4. Figure 2.4: The MuSE Architecture The MuSE architecture is as follows: Step 1: Generate Prompts The first stage is to leverage CuPL to create contextually relevant prompts for our multimodal model. CuPL is designed to generate prompts that effectively guide the model’s attention…
Figure 2.5
Figure 2.5. Figure 2.5: Classifying an image with CLIP at inference-time [Radford et al., 2021]. Baselines The baseline will be the standard method of CLIP image classification as in [PITH_FULL_IMAGE:figures/full_fig_p026_2_5.png]
Figure 2.6
Figure 2.6. Figure 2.6: A sample RESISC45 ground-truth airplane. 2.5.4 Fine-Grained Classification Query Image agaric mushroom Cosine Sim: 0.221 Cosine Sim: 0.201 Cosine Sim: 0.668 Cosine Sim: 0.795 [PITH_FULL_IMAGE:figures/full_fig_p032_2_6.png]
Figure 2.7
Figure 2.7. Figure 2.7: How MuSE Improves Fine-Grained Classification 32 [PITH_FULL_IMAGE:figures/full_fig_p032_2_7.png]
Figure 2.8
Figure 2.8. Figure 2.8: Ground Truth Bishop of Llandaff (left) compared to a generated image of the Bishop of Llandaff (right) 34 [PITH_FULL_IMAGE:figures/full_fig_p034_2_8.png]
Figure 2.9
Figure 2.9. Figure 2.9: Improved generated image of the Bishop of Llandaff For this image, our CuPL prompt was "Dahlia ’Bishop of Llandaff’ is a cultivar known for its dark red or maroon flowers and dark foliage." Even though it is similar in structure to the other CuPL prompts, it contains…
Figure 3.1
Figure 3.1. Figure 3.1: Images Generated with D3G for Race 4 40 [PITH_FULL_IMAGE:figures/full_fig_p040_3_1.png]
Figure 3.2
Figure 3.2. Figure 3.2: The image of Google Photos misclassifying Black people within the Photos application. When there are strong harmful demographic biases, these models can cause tremendous harms. One example was when Google Photo’s classified Black people as gorillas within a user’s al…
Figure 3.3
Figure 3.3. Figure 3.3: Racial compositions in face datasets Karkkainen and Joo [2021] 3.4 Methods We aim to create an ensemble of models to improve multimodal image classification accuracy, especially for models that are trained on data with a class imbalance. We test this method on standa…
Figure 3.4
Figure 3.4. Figure 3.4: The D3G Framework 3.4.1 Datasets For all of the results shown in this chapter, we classify images from the IdenProf test dataset. We selected this dataset because it provides a simple, applicable downstream task and because all of the images were collected and filter…
Figure 3.5
Figure 3.5. Figure 3.5: Classifying an image with CLIP at inference-time [Radford et al., 2021]. 3.5.2 Evaluation Breakdown We study three primary classification methods in this chapter. CLIP ViT-L/14 is our baseline, the standard method of multimodal classification as outlined in [PITH_FU…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Challenges and proposed solutions in modeling multimodal medical data: A systematic review

    cs.LG 2025-05 conditional novelty 4.0 of 10

    A systematic review of 69 medical multimodal modeling studies, organized by five reported challenges and the solution families proposed for each.

Reference graph

Works this paper leans on

29 extracted references · 15 canonical work pages · cited by 1 Pith paper

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    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 word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Leo Breiman. 1996. Bagging predictors. Machine learning, 24:123--140

    1996

  4. [4]

    Gong Cheng, Junwei Han, and Xiaoqiang Lu. 2017. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865--1883

    2017

  5. [5]

    Cimpoi, S

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. 2014. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition ( CVPR )

    work page 2014

  6. [6]

    J-P Dittrich and Bernhard Seeger. 2000. Data redundancy and duplicate detection in spatial join processing. In Proceedings of 16th International Conference on Data Engineering (Cat. No. 00CB37073), pages 535--546. IEEE

    work page 2000

  7. [7]

    Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. 2024. Datacomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 36

    2024

  8. [8]

    OT Jones, RN Matin, M Van der Schaar, K Prathivadi Bhayankaram, CKI Ranmuthu, MS Islam, D Behiyat, R Boscott, N Calanzani, Jon Emery, et al. 2022. Artificial intelligence and machine learning algorithms for early detection of skin cancer in community and primary care settings: a systematic review. The Lancet Digital Health, 4(6):e466--e476

    work page 2022

Show all 29 references
  1. [9]

    Kimmo Karkkainen and Jungseock Joo. 2021. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1548--1558

    2021

  2. [10]

    Wonjae Kim, Bokyung Son, and Ildoo Kim. 2021. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583--5594. PMLR

    2021

  3. [11]

    Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. 2023. Multimodal prompting with missing modalities for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14943--14952

    2023

  4. [12]

    Rushi Longadge and Snehalata Dongre. 2013. Class imbalance problem in data mining review. arXiv preprint arXiv:1305.1707

    2013 arXiv

  5. [13]

    S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. 2013. http://arxiv.org/abs/1306.5151 Fine-grained visual classification of aircraft . Technical report

    2013 arXiv

  6. [14]

    Helen Marsden, Caroline Morgan, Stephanie Austin, Claudia DeGiovanni, Marcello Venzi, Polychronis Kemos, Jack Greenhalgh, Dan Mullarkey, and Ioulios Palamaras. 2023. Effectiveness of an image analyzing ai-based digital health technology to identify non-melanoma skin cancer and...

    2023

  7. [15]

    National Institute of Standards and Technology . 2020. https://www.nist.gov/speech-testimony/facial-recognition-technology-frt-0 Facial recognition technology (frt) . Technical report, U.S. Department of Commerce, Washington, D.C

    2020

  8. [16]

    Maria-Elena Nilsback and Andrew Zisserman. 2008. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing

    2008

  9. [17]

    Noseong Park, Mahmoud Mohammadi, Kshitij Gorde, Sushil Jajodia, Hongkyu Park, and Youngmin Kim. 2018. Data synthesis based on generative adversarial networks. arXiv preprint arXiv:1806.03384

    2018 arXiv

  10. [18]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M \"u ller, Joe Penna, and Robin Rombach. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952

    2023 arXiv

  11. [19]

    Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. 2023. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691--15701

    2023

  12. [20]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  13. [21]

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821--8831. Pmlr

    2021

  14. [22]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684--10695

    2022

  15. [23]

    Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. 2022. Sus-x: Training-free name-only transfer of vision-language models. arXiv preprint arXiv:2211.16198

    2022 arXiv

  16. [24]

    Matthew Wallingford, Vivek Ramanujan, Alex Fang, Aditya Kusupati, Roozbeh Mottaghi, Aniruddha Kembhavi, Ludwig Schmidt, and Ali Farhadi. 2024. Neural priming for sample-efficient adaptation. Advances in Neural Information Processing Systems, 36

    2024

  17. [25]

    Zilong Wang, Zhaohong Wan, and Xiaojun Wan. 2020. Transmodality: An end2end fusion method with transformer for multimodal sentiment analysis. In Proceedings of The Web Conference 2020, pages 2514--2520

    2020

  18. [26]

    David Wen, Saad M Khan, Antonio Ji Xu, Hussein Ibrahim, Luke Smith, Jose Caballero, Luis Zepeda, Carlos de Blas Perez, Alastair K Denniston, Xiaoxuan Liu, et al. 2022. Characteristics of publicly available skin cancer image datasets: a systematic review. The Lancet Digital Hea...

    2022

  19. [27]

    Hui Xiong, Gaurav Pandey, Michael Steinbach, and Vipin Kumar. 2006. Enhancing data analysis with noise removal. IEEE transactions on knowledge and data engineering, 18(3):304--319

    2006

  20. [28]

    Ze Yang, Tiange Luo, Dong Wang, Zhiqiang Hu, Jun Gao, and Liwei Wang. 2018. Learning to navigate for fine-grained classification. In Proceedings of the European conference on computer vision (ECCV), pages 420--435

    2018

  21. [29]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence '...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.