Pith. sign in

REVIEW 6 major objections 7 minor 51 references

Vision-Language Model Selection and Reuse for Downstream Adaptation

T0 review · 6 major / 7 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 Model Label Learning pre-labels each vision-language model once on a semantic graph, then matches task captions to graph labels and ensembles top models per class, beating ImageNet and text-only selection on 17 datasets.

desk verdict A genuine benchmark and a plausible per-class VLM selection pipeline, but the evaluation is tuned on the test set and the key transfer in Eq. (6) isn't validated. read the letter →

arxiv 2501.18271 v2 pith:MDD55337 submitted 2025-01-30 cs.LG cs.AI

Hao-Zhe Tan , Zhi Zhou , Yu-Feng Li , Lan-Zhe Guo This is my paper · ORCID
classification cs.LGcs.AI
keywords vision-languagemodelszero-shotclassificationmodelselectionreusehubsemanticgraphensemblepredictionLabelLearning
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

This paper tries to solve a practical bottleneck: with hundreds of open-source vision-language models available, no single model wins every task, yet testing all of them on a new task costs too much. It proposes Model Label Learning (MLL), in which each candidate model is pre-tested once on a fixed semantic graph of visual concepts to produce a 'label' describing its strengths. When a new task arrives, the system matches the task's classes to graph nodes using text captions and an embedding similarity, predicts each model's per-class accuracy from the labels, and ensembles the top models per class. The authors run this on 49 models and 17 datasets and report higher average accuracy than the ImageNet baseline and the previous text-only selector, at a fraction of the evaluation cost. If the transfer assumption holds, model hubs could serve new tasks instantly from precomputed labels.

What carries the argument

The load-bearing object is the semantic graph $G$: more than 9,000 visual-concept nodes, each carrying sample images and a text caption formed as '{concept} which is {definition}', with edges from a lexical database's hypernym/hyponym relations. Each VLM is pre-tested on these nodes to produce label $S_m$ of similarity scores $s^v_{m,x}$ and per-node precision $p_{m,v}$ (Eq. 5). For a new task, class captions are generated with a large language model and embedded; the transfer matrix $Z=(z_{vy})$ holds the cosine similarities between the top-$k$ matched graph-node captions and task-class captions. The identity that does the work is Eq. (6), $p_{m,y} = \sum_v p_{m,v} z_{vy}$, which turns graph-measured strengths into a predicted per-class accuracy for the target task, then Eq. (7) blends it with overall performance and Eqs. (8)-(10) build class-wise ensembles.

What would settle it

On any of the 17 benchmark target tasks with per-class ground truth, compute the proposed $p_{m,y}$ from Eq. (6) for each class and compare it with the model's true per-class zero-shot accuracy. If a class whose matched semantic nodes carry high $p_{m,v}$ still shows near-chance true accuracy on that dataset—for example a medical-image class matched to natural-picture concepts—the transfer assumption in Eq. (6) fails.

Watch

Extended reading notes

Core claim

The central claim is that a VLM's per-class accuracy on an arbitrary downstream task can be predicted from its measured per-class accuracies on a fixed semantic graph, without running the model on the task. The predicted accuracy for class $y$ is $p_{m,y} = \sum_{v} p_{m,v} z_{vy}$, where $p_{m,v}$ is the model's precision on graph node $v$ measured on sample images, and $z_{vy}$ is the text-embedding similarity between the node's caption and the task class's caption. Blending this per-class prediction with the model's overall predicted accuracy gives a reuse metric, and the top-$k$ models per class are ensembled with entropy-based weights. On a benchmark of 49 models and 17 tasks, the paper reports higher average accuracy than selecting the best ImageNet model or using a prior text-only selection method.

Load-bearing premise

The load-bearing premise is that per-class accuracies measured on natural-photo sample images at matched semantic concepts transfer to arbitrary downstream classes through text-caption similarity, so if a task's images or class descriptions differ from those samples, the predicted per-class accuracy can be wrong.

Editorial extensions

If this is right

  • A model hub can pre-test every incoming VLM once on the semantic graph, so a new downstream task is served by caption matching and ensembling without running any candidate on the task's images.
  • Hub growth improves selection: as more VLMs are added, per-class ensembling has better candidates to draw from, raising average downstream accuracy.
  • Class-level reuse works: picking a different top-$k$ set per class, weighted by prediction entropy, outperforms selecting a single global best model on most of the 17 tasks.
  • The method beats the two baselines on average: ImageNet-based selection and the prior text-only selector both trail the proposed reuse metric on average accuracy.

Reading between the lines

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

  • The labeling step could be reused for other model families if their outputs can be scored on the same graph; for example, generative image models could be labeled by how faithfully they reproduce each concept, giving the same 'serve a new task without evaluation' property.
  • Selection quality inherits the caption generator's blind spots: a cheaper or domain-specific captioner may change the transfer matrix and thus the chosen models, which is a direct testable extension.
  • The per-class routing design implies the hub can assign different experts to different classes of one task, not just pick one model per task; extending the benchmark to per-class accuracy would expose where that routing helps most.
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

6 major / 7 minor

Summary. The paper proposes Model Label Learning (MLL), a paradigm for selecting and reusing pre-trained vision-language models (VLMs) for downstream zero-shot tasks without running all candidate models on the target data. In the model-labeling module, each VLM is pre-tested on a WordNet-based semantic graph of 9,055 nodes whose images come from ImageNet-family datasets; this yields per-node similarity scores. In model selection, GPT-4 captions of the target task classes are matched to graph-node captions via text-embedding cosine similarity, and Eq. (6) converts per-node precision estimates into per-class predictions. In model reuse, the top-k models per class are ensembled using entropy-weighted confidence (Eqs. 8–10). The authors introduce a benchmark of 49 VLMs and 17 target datasets and report average accuracy improvements over an ImageNet baseline (INB) and ModelGPT for k=1 and k=3.

Significance. The central claim is substantively interesting: if the transfer in Eq. (6) holds, model hubs could pre-test VLMs once on a semantic graph and serve new tasks by caption matching and ensembling, avoiding per-task evaluation. The paper contributes a reusable benchmark, a public code release, and an explicit limitation statement in the conclusion. However, the evidence is currently thin in several load-bearing places: the transfer assumption in Eq. (6) is not validated against per-class ground truth, the ensemble entropy weighting in Eq. (9) is underspecified, hyperparameters α and k are tuned on the same evaluation benchmark, and no error bars or significance tests accompany the reported averages. The claimed selection capability is thus not yet cleanly separated from the gains of ensembling and reuse.

major comments (6)
  1. [§4.2, Eq. (6)] The central transfer assumption is unvalidated. Eq. (6) assumes that per-class accuracy on an arbitrary target task, pm,y = Σ_v pm,v · zvy, can be predicted by combining ImageNet-family per-synset accuracies with caption-embedding similarities. The paper only reports average accuracy over 17 tasks, never comparing predicted pm,y to actual per-class accuracy, nor evaluating whether the induced model ranking per class is better than random. Since the final pipeline also includes a per-class ensemble (§4.3), the observed average gains could be carried by the ensembling stage rather than by selection. Please add per-class validation (e.g., correlation or rank metrics between predicted and actual per-class accuracy) and an ablation that isolates the selection component, such as comparing against oracle per-class selection.
  2. [§4.1–4.2, Eq. (5)] The 'target-task-independent' labeling claim is internally inconsistent with the definition of precision pm,v in Eq. (5), which computes the argmax over V_Selected, a target-task-specific set of matched nodes chosen at identification time (Algorithm 1, steps 2–3). Thus pm,v is not a fixed label produced in the submission stage; it is recomputed for each downstream task. The submission stage only produces raw similarity scores Sm (Eq. 4), and the precision is a query-time quantity. Please clarify what exactly constitutes the 'model label' and revise the claims in §1 and §4.4 that the labeling process is completed when the model is uploaded.
  3. [§4.3, Eq. (9)] The entropy weighting is underspecified and the accompanying text contradicts the formula. H is applied to the set of raw cosine similarities {sim(Im(x), Tm(y)) | y ∈ YT}, but entropy is normally defined on a probability distribution; raw cosine similarities are not normalized and may be negative. The text states that the method 'assign[s] lower weights to models with high confidence when they are overconfident,' but Eq. (9) gives higher weights to high-entropy (low-confidence) models, with no conditional on overconfidence. Please state the exact entropy definition (e.g., entropy of the softmax probabilities in Eq. (8), with temperature τ included) and reconcile the description with the formula.
  4. [§5.2, Tables 3–4] The hyperparameters α and k are selected by sweeping on the same 17-task benchmark used for the final comparison (α = 0.7 in §5.2; k = 1 and 3 in Tables 1–2). This amounts to tuning on the test set and makes the reported average gains difficult to interpret. Moreover, no error bars, standard deviations, or significance tests are reported for any table or Figure 3. Please provide a validation-based hyperparameter selection, or at least a sensitivity analysis with variance over model-hub subsamples or repeated randomized runs, so that the improvement over INB/ModelGPT can be assessed statistically.
  5. [§4.1 and §5.1] The construction of the semantic graph is underspecified for synsets outside the ImageNet-1000 label set. The paper states that 9,055 WordNet synsets are used and images are randomly drawn from five ImageNet-family datasets, but it does not explain how images are assigned to synsets that are not among the ImageNet classes (the sample datasets cover at most 1,000 classes each, and some cover only 200). This matters because Eq. (5) requires a non-empty Xv for every node, and Eq. (6) inherits any noise in those per-node image sets. Please describe the image-assignment procedure (e.g., matching by synset labels, using hypernym/hyponym relations, or manual curation) and report coverage statistics, such as the number of nodes with at least one image and the average |Xv|.
  6. [§1 and §5.1] The paper advertises that 'ground-truth model ranking for each target task is provided for evaluation,' yet the experiments never evaluate ranking quality (e.g., Spearman/Kendall correlation between predicted and true rankings, or top-k recall). Tables 1–2 only report downstream accuracy of the selected/reused model(s). Since the paper frames the problem as model selection, please add selection-quality metrics; otherwise the 'selecting' half of the central claim is only indirectly supported by end-to-end accuracy.
minor comments (7)
  1. [§2, §5.2, Tables 1–2] The method cited as 'Zohar et al. (2023)' is called LOVM in §2 but ModelGPT in §5.2 and the tables. Since ModelGPT is a distinct published method, please align the naming with the cited paper and verify that the comparison is to the correct baseline.
  2. [§4.3, Eq. (8)] Eq. (8) omits the temperature τ that appears in the zero-shot prediction defined in Eq. (1). Please clarify whether the ensemble uses each model's native temperature or a fixed value, since the entropy weighting in Eq. (9) depends on the resulting probabilities.
  3. [Table 3] The α sweep reports only average accuracy over 17 tasks. Per-task results would help assess whether the robustness claim holds across different domains, especially for tasks where the proposal already underperforms the baseline.
  4. [Figure 3] The scalability plot aggregates 30 randomly generated expansion schemes but shows no error bars or dispersion. Please report standard deviations or quartiles to support the claim that performance grows reliably with hub size.
  5. [§4.2] The transfer matrix Z is described as containing zvy = similarity when v is among the top-k nodes for class y and 0 otherwise, but the text does not specify whether the nonzero entries are used as raw cosine similarities or normalized (e.g., row-wise softmax). This affects the scale of pm,y in Eq. (6).
  6. [Conclusion] The limitation statement in the conclusion only mentions the focus on VLMs and visual classification tasks. Please also acknowledge the lack of per-class validation of Eq. (6) and the on-benchmark tuning of α and k, since these are more immediate limitations of the current evidence.
  7. [Throughout] Minor language issues include 'count k of reused models pre-class' in Algorithm 1 (should be 'per class') and 'predication' in §5.1 (should be 'prediction').

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: Eq. (6) is a transparent transfer heuristic, not a fitted prediction; self-citations are contextual related work.

full rationale

The claimed derivation chain—pre-test VLMs on WordNet-based semantic graph images (Sec. 4.1), match target classes to graph nodes by caption similarity, form pm,y = sum_v pm,v * zvy (Eq. 6), rank models by rm,y (Eq. 7), and ensemble top-k—does not use target-task labels or ground-truth rankings in fitting. Eq. (6) is an explicit inductive transfer rule: pm,v is measured on sample datasets, zvy is a text-embedding similarity, and pm,y is the weighted estimate. It is not a restatement of the target accuracy; target accuracy is later measured and reported in Tables 1 and 2. No parameter is fit to target data and then renamed as a prediction; alpha = 0.7 and k are hyperparameters chosen from ablations on the same benchmark, which is a methodological weakness (risk of selection bias) but not circularity, because those ablations are not part of the derivation chain. The Learnware citations in Sec. 2 are related-work context and are not invoked to justify any premise or to rule out alternatives; the paper explicitly distinguishes MLL from learnware because learnware requires training data. The WordNet/GPT-4 caption-matching assumption in Eq. (6) is a correctness risk—the paper does not validate per-class transfer against ground truth—but an unvalidated assumption is not equivalent by construction to its output. Overall, the central mechanism is self-contained and externally benchmarked, so no circular step is present.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the ability of a WordNet-based semantic graph with ImageNet-family sample images to produce per-class accuracy estimates that transfer to unseen tasks via text-embedding matching. The main tunable inputs are alpha, the number of reused models k, the number of matched nodes per class, and the graph/image sampling choices; none of these are grounded in theory or a held-out split.

free parameters (5)
  • alpha = 0.7
    Balances per-class and overall performance in Eq. 7; chosen after ablation on the 17 benchmark tasks (Table 3), not on a held-out split.
  • k (reused models per class) = 1 and 3
    Number of models ensembled per class; results in Tables 1-2 and 4 show k=3 yields only a slight average gain, and k is selected post hoc on the same benchmark.
  • top-k matched semantic nodes per class = not reported
    Number of graph nodes retained in transfer matrix Z (Eq. 6) is not specified in the implementation details; it affects the estimated per-class scores.
  • images per semantic node = up to 75
    Number of sample images used to compute each node label (Section 5.2); chosen by the authors, affects label stability.
  • semantic graph size = 9055 WordNet synsets
    Number of nodes in the semantic graph; arbitrary but large, affects coverage and compute.
assumptions (6)
  • domain assumption WordNet synsets and their hypernym/hyponym relationships form a valid semantic graph for visual concepts.
    Section 4.1: the graph is used as the evaluation dataset for pre-testing all VLMs.
  • domain assumption ImageNet-family sample datasets provide representative images for the visual concepts in WordNet synsets.
    Section 5.1: if these images are not representative, model labels misestimate per-concept performance.
  • domain assumption GPT-4 generated captions for target classes accurately describe the visual content of those classes.
    Section 4.2: if captions are off, the transfer matrix Z misaligns classes to nodes.
  • domain assumption Cosine similarity between text embeddings of task captions and synset captions reliably indicates semantic correspondence.
    Section 4.2: the entire transfer depends on this similarity.
  • ad hoc to paper The linear transfer pm,y = sum_v pm,v * zvy (Eq. 6) is a valid estimator of per-class accuracy on the target task.
    Introduced without theoretical justification and not validated against per-class ground truth.
  • domain assumption Entropy weighting in Eq. (9) reduces the influence of overconfident models and improves ensemble accuracy.
    Asserted based on Vishniakov et al. (2024), but Eq. 9 as written applies entropy to raw similarities rather than probabilities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vision-Language Model Selection and Reuse for Downstream Adaptation." pith.science (2026). https://pith.science/paper/MDD55337

@misc{pith2026250118271,
  author       = {Pith},
  title        = {Pith review of: Vision-Language Model Selection and Reuse for Downstream Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MDD55337}},
  note         = {Machine review of arXiv:2501.18271}
}
read the original abstract

Pre-trained Vision-Language Models (VLMs) are becoming increasingly popular across various visual tasks, and several open-sourced VLM variants have been released. However, selecting the best-performing pre-trained VLM for a specific downstream task is challenging since no single VLM can achieve promising performance on all downstream tasks, and evaluating all available VLMs is impossible due to time and data limitations. To address this problem, this paper proposes a novel paradigm to select and reuse VLM for downstream tasks, called Model Label Learning (MLL). The proposal contains three key modules: \emph{model labeling}, which assigns labels to each VLM to describe their specialty and utility; \emph{model selection}, which matches the requirements of the target task with model labels; and \emph{model reuse}, which applies selected VLMs to the target task in an ensemble manner. The proposal is highly computationally efficient and growable since the model labeling process is completed target task independent and the ability could grow with the number of candidate VLMs. We also introduce a new benchmark for evaluating VLM selection methods, including 49 VLMs and 17 target task datasets. Experimental results clearly demonstrate the effectiveness of the proposed method for selecting and reusing VLMs.

Figures

Figures reproduced from arXiv: 2501.18271 by the authors.

Figure 1
Figure 1. The spider charts measure 49 models’ capabilities across 7 downstream tasks and classes within a task, showing that the best-performing models vary across downstream tasks and classes, highlighting the importance of model selection for VLM. The evaluated 49 models align with those in the model hub, as discussed in Section 5.1. selection since they only handle either image or text output and cannot incorporate data f… view at source ↗
Figure 2
Figure 2. The framework of MLL paradigm. Models added to the hub first undergo a pre-testing phase, during which they are assigned labels that describe their specific functionalities in the labeling module. When a downstream task is presented, the system selects relevant models in the selection module and ensembles them to address the task. the VLM fm. There are two stages in our setting: sub￾mission stage for model developer… view at source ↗
Figure 3
Figure 3. Average performance on 17 downstream tasks with the scaling of the model hub. Implementation Details. We adopt the official code to im￾plement ModelGPT. For a fair comparison, the experiment utilizes the ground-truth performance of VLMs on Sample Datasets for ModelGPT to train its linear model, and then evaluate it on the benchmark. For both INB and ModelGPT, the experiment selects the model with the highest predict… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 41 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Learning multiple layers of features from tiny images

    Alex, K. Learning multiple layers of features from tiny images. 2009

    work page 2009

  3. [3]

    Food-101--mining discriminative components with random forests

    Bossard, L., Guillaumin, M., and Van Gool, L. Food-101--mining discriminative components with random forests. In Proceedings of the 13th European Conference on Computer Vision, pp.\ 446--461, 2014

    work page 2014

  4. [4]

    Remote sensing image scene classification: Benchmark and state of the art

    Cheng, G., Han, J., and Lu, X. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105 0 (10): 0 1865--1883, 2017

    work page 2017

  5. [5]

    Describing textures in the wild

    Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., and Vedaldi, A. Describing textures in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 3606--3613, 2014

    2014

  6. [6]

    An analysis of single-layer networks in unsupervised feature learning

    Coates, A., Ng, A., and Lee, H. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the 14th International Conference on Artificial Intelligence and Statistics, pp.\ 215--223, 2011

    work page 2011

  7. [7]

    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 Proceedings of 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 248--255, 2009

    work page 2009

  8. [8]

    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., Uszkoreit, J., and Houlsby, N. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the 9th International Conference on Learning Representations, 2021

    work page 2021

Show all 51 references
  1. [9]

    Data determines distributional robustness in Contrastive Language Image Pre-training (CLIP)

    Fang, A., Ilharco, G., Wortsman, M., Wan, Y., Shankar, V., Dave, A., and Schmidt, L. Data determines distributional robustness in Contrastive Language Image Pre-training (CLIP) . In Proceedings of the 39th International Conference on Machine Learning, pp.\ 6216--6234, 2022

    2022

  2. [10]

    EVA : Exploring the limits of masked visual representation learning at scale

    Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., and Cao, Y. EVA : Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 19358--19369, 2023

    2023

  3. [11]

    J., Erhan, D., Carrier, P

    Goodfellow, I. J., Erhan, D., Carrier, P. L., Courville, A., Mirza, M., Hamner, B., Cukierski, W., Tang, Y., Thaler, D., Lee, D.-H., et al. Challenges in representation learning: A report on three machine learning contests. In Proceedings of the 20th International Conference o...

    2013

  4. [12]

    Identifying useful learnwares for heterogeneous label spaces

    Guo, L.-Z., Zhou, Z., Li, Y.-F., and Zhou, Z.-H. Identifying useful learnwares for heterogeneous label spaces. In Proceedings of the 40th International Conference on Machine Learning, pp.\ 12122--12131, 2023

    2023

  5. [13]

    Robust semi-supervised learning in open environments

    Guo, L.-Z., Jia, L.-H., Shao, J.-J., and Li, Y.-F. Robust semi-supervised learning in open environments. Frontiers of Computer Science, 19 0 (8): 0 198345, 2025

    2025

  6. [14]

    The many faces of robustness: A critical analysis of out-of-distribution generalization

    Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Compute...

    2021

  7. [15]

    Natural adversarial examples

    Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., and Song, D. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 15262--15271, 2021 b

    2021

  8. [16]

    OpenCLIP , 2021

    Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., and Schmidt, L. OpenCLIP , 2021

    2021

  9. [17]

    Scaling up visual and vision-language representation learning with noisy text supervision

    Jia, C., Yang, Y., Xia, Y., Chen, Y.-T., Parekh, Z., Pham, H., Le, Q., Sung, Y.-H., Li, Z., and Duerig, T. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, pp.\ 4...

    2021

  10. [18]

    CLEVR : A diagnostic dataset for compositional language and elementary visual reasoning

    Johnson, J., Hariharan, B., Van Der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., and Girshick, R. CLEVR : A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 2...

    2017

  11. [19]

    3D object representations for fine-grained categorization

    Krause, J., Stark, M., Deng, J., and Fei-Fei, L. 3D object representations for fine-grained categorization. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pp.\ 554--561, 2013

    2013

  12. [20]

    Gradient-based learning applied to document recognition

    LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86 0 (11): 0 2278--2324, 1998

    1998

  13. [21]

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

    Li, J., Li, D., Xiong, C., and Hoi, S. BLIP : Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pp.\ 12888--12900, 2022

    2022

  14. [22]

    Towards safe weakly supervised learning

    Li, Y.-F., Guo, L.-Z., and Zhou, Z.-H. Towards safe weakly supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43 0 (1): 0 334--346, 2021

    2021

  15. [23]

    A ConvNet for the 2020s

    Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., and Xie, S. A ConvNet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11976--11986, 2022

    2022

  16. [24]

    Foundation model is efficient multimodal multitask model selector

    Meng, F., Shao, W., Peng, Z., Jiang, C., Zhang, K., Qiao, Y., and Luo, P. Foundation model is efficient multimodal multitask model selector. In Advances in Neural Information Processing Systems, pp.\ 33065--33094, 2023

    2023

  17. [25]

    Miller, G. A. WordNet : A lexical database for English . Communications of the ACM, 38 0 (11): 0 39–41, 1995

    1995

  18. [26]

    LEEP : A new measure to evaluate transferability of learned representations

    Nguyen, C., Hassner, T., Seeger, M., and Archambeau, C. LEEP : A new measure to evaluate transferability of learned representations. In Proceedings of the 37th International Conference on Machine Learning, pp.\ 7294--7305, 2020

    2020

  19. [27]

    and Zisserman, A

    Nilsback, M.-E. and Zisserman, A. Automated flower classification over a large number of classes. In Proceedings of 2008 6th Indian Conference on Computer Vision, Graphics & Image Processing, pp.\ 722--729, 2008

    2008

  20. [28]

    GPT-4 technical report

    OpenAI. GPT-4 technical report. CoRR, abs/2303.08774, 2023

    2023 arXiv

  21. [29]

    A., Ahmed, N., and Faloutsos, C

    Park, N., Rossi, R. A., Ahmed, N., and Faloutsos, C. MetaGL : Evaluation-free selection of graph learning models via meta-learning. In Proceedings of the 11th International Conference on Learning Representations, 2023

    2023

  22. [30]

    M., Vedaldi, A., Zisserman, A., and Jawahar, C

    Parkhi, O. M., Vedaldi, A., Zisserman, A., and Jawahar, C. Cats and dogs. In Proceedings of 2012 IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 3498--3505, 2012

    2012

  23. [31]

    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 Proceedings of the 38th International Conference on Machine Learning, pp.\ 8...

    2021

  24. [32]

    Do ImageNet classifiers generalize to ImageNet ? In Proceedings of the 36th International Conference on Machine Learning, pp.\ 5389--5400, 2019

    Recht, B., Roelofs, R., Schmidt, L., and Shankar, V. Do ImageNet classifiers generalize to ImageNet ? In Proceedings of the 36th International Conference on Machine Learning, pp.\ 5389--5400, 2019

    2019

  25. [33]

    LAION-5B : An open large-scale dataset for training next generation image-text models

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al. LAION-5B : An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, p...

    2022

  26. [34]

    R., and Shah, M

    Soomro, K., Zamir, A. R., and Shah, M. UCF101 : A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012

    2012 arXiv

  27. [35]

    WIT : Wikipedia-based image text dataset for multimodal multilingual machine learning

    Srinivasan, K., Raman, K., Chen, J., Bendersky, M., and Najork, M. WIT : Wikipedia-based image text dataset for multimodal multilingual machine learning. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp.\ 24...

    2021

  28. [36]

    Stallkamp, J., Schlipsing, M., Salmen, J., and Igel, C. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. Neural Networks, 32: 0 323--332, 2012

    2012

  29. [37]

    Handling learnwares developed from heterogeneous feature spaces without auxiliary data

    Tan, P., Tan, Z.-H., Jiang, Y., and Zhou, Z.-H. Handling learnwares developed from heterogeneous feature spaces without auxiliary data. In Proceedings of the 32nd International Joint Conference on Artificial Intelligence, pp.\ 4235--4243, 2023

    2023

  30. [38]

    Beimingwu: A learnware dock system

    Tan, Z.-H., Liu, J.-D., Bi, X.-D., Tan, P., Zheng, Q.-C., Liu, H.-T., Xie, Y., Zou, X.-C., Yu, Y., and Zhou, Z.-H. Beimingwu: A learnware dock system. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 5773--5782, 2024

    2024

  31. [39]

    T., Nguyen, C

    Tran, A. T., Nguyen, C. V., and Hassner, T. Transferability and hardness of supervised classification tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 1395--1405, 2019

    2019

  32. [40]

    S., Linmans, J., Winkens, J., Cohen, T., and Welling, M

    Veeling, B. S., Linmans, J., Winkens, J., Cohen, T., and Welling, M. Rotation equivariant CNNs for digital pathology. In Proceedings of the 21st International Conference on Medical Image Computing and Computer Assisted Intervention, pp.\ 210--218, 2018

    2018

  33. [41]

    ConvNet vs Transformer , Supervised vs CLIP : Beyond ImageNet accuracy

    Vishniakov, K., Shen, Z., and Liu, Z. ConvNet vs Transformer , Supervised vs CLIP : Beyond ImageNet accuracy. In Proceedings of the 41st International Conference on Machine Learning, 2024

    2024

  34. [42]

    Wang, H., Ge, S., Lipton, Z., and Xing, E. P. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, 2019

    2019

  35. [43]

    Transformers: State-of-the-art natural language processing

    Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al. Transformers: State-of-the-art natural language processing. In Proceedings of 2020 Conference on Empirical Methods in Natural Language Processing, pp.\ 38...

    2020

  36. [44]

    Model reuse with reduced kernel mean embedding specification

    Wu, X.-Z., Xu, W., Liu, S., and Zhou, Z.-H. Model reuse with reduced kernel mean embedding specification. IEEE Transactions on Knowledge and Data Engineering, 35 0 (1): 0 699--710, 2021

    2021

  37. [45]

    LogME : Practical assessment of pre-trained models for transfer learning

    You, K., Liu, Y., Wang, J., and Long, M. LogME : Practical assessment of pre-trained models for transfer learning. In Proceedings of the 38th International Conference on Machine Learning, pp.\ 12133--12143, 2021

    2021

  38. [46]

    CoCa : Contrastive captioners are image-text foundation models

    Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., and Wu, Y. CoCa : Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022

    2022

  39. [47]

    S., Neumann, M., Dosovitskiy, A., et al

    Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A. S., Neumann, M., Dosovitskiy, A., et al. A large-scale study of representation learning with the visual task adaptation benchmark. CoRR, abs/1910.04867, 2019

    1910 arXiv

  40. [48]

    Model spider: Learning to rank pre-trained models efficiently

    Zhang, Y.-K., Huang, T.-J., Ding, Y.-X., Zhan, D.-C., and Ye, H.-J. Model spider: Learning to rank pre-trained models efficiently. Advances in Neural Information Processing Systems, pp.\ 13692--13719, 2024

    2024

  41. [49]

    You only submit one image to find the most suitable generative model

    Zhou, Z., Guo, L.-Z., Song, P.-X., and Li, Y.-F. You only submit one image to find the most suitable generative model. In Proceedings of the 37th Annual Conference on Neural Information Processing Systems Workshop on Diffusion Models, 2023

    2023

  42. [50]

    and Tan, Z.-H

    Zhou, Z.-H. and Tan, Z.-H. Learnware: Small models do big. CoRR, abs/2210.03647, 2022

    2022 arXiv

  43. [51]

    LOVM : Language-only vision model selection

    Zohar, O., Huang, S.-C., Wang, K.-C., and Yeung, S. LOVM : Language-only vision model selection. In Advances in Neural Information Processing Systems, pp.\ 33120--33132, 2023

    2023

Pith tools

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