Hierarchical Multi-Positive Contrastive Learning for Patent Image Retrieval

Angelos Nalmpantis; Emma Beauxis-Aussalet; Kshitij Kavimandan; Robert-Jan Sips

REVIEW 3 major objections 3 minor 26 references

Hierarchical Multi-Positive Contrastive Learning for Patent Image Retrieval

T0 review · 3 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper proposes replacing the standard single-positive contrastive loss with a hierarchy-weighted multi-positive loss, using the Locarno International Classification to assign patent images multiple positive pairs with decreasing…

desk verdict Clean, honest core idea with a misleading abstract: patent-level mAP drops in the paper's own tables contradict the 'enhances retrieval' headline, but the method is a reasonable baseline and worth a referee. read the letter →

arxiv 2506.13496 v3 pith:G4TKDPWJ submitted 2025-06-16 cs.CV cs.IRcs.LG

Kshitij Kavimandan , Angelos Nalmpantis , Emma Beauxis-Aussalet , Robert-Jan Sips This is my paper · ORCID

classification cs.CVcs.IRcs.LG

keywords patentimageretrievalhierarchicalcontrastivelearningmulti-positivelossLocarnoInternationalClassificationtechnicaldrawinglow-parametermodelsvisiontransformermultimodalCLIPfine-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

Patent images are technical drawings whose retrieval rankings often ignore how patents are categorized. The paper argues that the Locarno International Classification (LIC) hierarchy encodes useful visual-semantic closeness: two images from the same patent should be pulled together most strongly, images from the same LIC subclass less so, and images from the same main class least but still positively. The proposed hierarchical multi-positive contrastive loss replaces the single-positive contrastive loss with a weighted sum over many positives, and the reported experiments on DeepPatent2 show that this raises retrieval accuracy at subclass and main class levels across ResNet, ViT, and CLIP encoders. A reader should care because the method is simple enough for low-parameter models and plugs into existing fine-tuning pipelines without a new architecture.

What carries the argument

The load-bearing object is the hierarchical multi-positive contrastive loss of Equation 4, built from the relevance function $h_{ij}$ of Equation 2 and its normalization $H_i$ of Equation 3. $h_{ij}$ converts the LIC taxonomy into training weights, so the loss maximizes the cosine similarity of an anchor to several positives, weighted by how close they are in the patent hierarchy; the standard contrastive loss of Equation 1 is the special case with a single positive pair scored 1. This mechanism is what injects the hierarchy into the embedding space, and it generalizes to any taxonomy by redefining $h_{ij}$.

What would settle it

Retrain the same encoders with the same batches but shuffled or random LIC labels; if subclass- and main-class mAP still rises above conventional contrastive learning, the gains are not caused by the taxonomy's semantic structure. A second check is to count, for a fixed query, how often the top retrieved same-subclass images are visually unrelated drawings: enough such cases would indicate the loss is aligning administrative labels rather than visual content.

Watch

Extended reading notes

Core claim

The central claim is that replacing the standard contrastive loss with a hierarchy-weighted multi-positive version improves patent image retrieval. Concretely, for an anchor image $x_i$, every other image $\tilde{x}_j$ in the batch receives a relevance score $h_{ij}$: $s_p=1$ for same patent ID, $s_s=0.35$ for same LIC subclass, $s_m=0.2$ for same main class, and $0$ otherwise, and the loss in Equation 4 uses $h_{ij}/H_i$ as weights over the positive log-probabilities. The authors report that this formulation outperforms conventional contrastive fine-tuning at the subclass and main class levels on almost every model tested, with the largest gains on small models such as ResNet-18 and ViT-Tiny, while acknowledging a small patent-ID level drop for larger ViT and CLIP models. The hierarchy signal also produces more separable subclass clusters in the embedding space, as shown by PCA projections.

Load-bearing premise

The method assumes the LIC taxonomy tracks visual relevance: images sharing a subclass or main class are treated as usefully similar even when they depict different objects, so the hierarchy can serve as a training signal.

Editorial extensions

If this is right

At the subclass and main class levels, retrieval quality improves on the DeepPatent2 test set for ResNet, ViT, and CLIP encoders when the hierarchical loss replaces standard contrastive fine-tuning.
Small models, including ResNet-18 and ViT-Tiny, receive the largest relative gains, which suggests useful patent retrieval can run on hardware with limited compute.
The method trades a small amount of exact patent-ID precision on larger ViT and CLIP models for broader-level ranking gains, an explicit design trade-off.
Because Equation 4 only depends on the scoring function $h_{ij}$, the same recipe transfers to other classification systems, such as the Cooperative Patent Classification.
Adding the text-supervision term of Equation 5 further improves multimodal CLIP retrieval, so the hierarchy signal composes with language supervision.

Reading between the lines

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

A natural experiment the paper does not run is to permute the LIC labels during training; if subclass-level mAP gains persist under shuffled taxonomy labels, the effect would come from extra positives in general rather than from hierarchy semantics.
The hand-set weights $s_p=1$, $s_s=0.35$, $s_m=0.2$ are arbitrary; optimizing them per level or per subclass may widen or narrow the reported trade-off, and the paper's own claim would be strengthened by a sensitivity analysis.
The larger gains for small models suggest the hierarchy acts partly as a regularizer; one could test this by measuring whether HMCL also improves linear probing on held-out subclasses not seen during fine-tuning.

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

Quick take: this is a tidy little workshop paper with a honest core idea and unusually broad experiments for six pages, but the abstract says 'enhances the retrieval results' without qualification, and the tables don't support that at the patent level for five model variants.

What's genuinely new: the weighted multi-positive contrastive loss over LIC levels, Eq. 4 with h_ij defined in Eq. 2. It's a straightforward extension of soft-label contrastive learning, but I haven't seen it applied to patent retrieval, and the graded hierarchy (patent > subclass > main class) is a reasonable inductive bias. The authors also correctly note that Eq. 1 is a special case, and they run ResNet, ViT, CLIP vision-only, and CLIP vision+language with multiple seeds. The standard deviations are small and roughly consistent across methods, and the paper is honest in Section 4 that higher-level gains come with a patent-level cost on larger models.

The soft spots are real but not disqualifying. First, the abstract and conclusion claim gains at 'all hierarchical levels,' which contradicts the patent-ID mAP drops for ViT-Small, ViT-Base, ViT-Large, CLIP-B/16*, CLIP-L/14*, CLIP-B/16, and CLIP-L/14 (Tables 1 and 2). The body does acknowledge this trade-off in one sentence, so the mismatch is mostly an abstract problem, but it's still a misleading headline. Second, the weights sp=1, ss=0.35, sm=0.2 and lambda=0.2 are set by hand with no sensitivity analysis or ablation. The authors say these values 'offer a balanced performance,' but that's an empirical claim without support. Third, there's no comparison with other hierarchical or soft-label contrastive methods, so it's hard to know if the specific weighting matters or just the multi-positive extension. Fourth, no code is released; the training details are sufficient for replication, but only barely.

On the stress-test note: yes, the abstract overclaims, and a label-shuffle control would have separated the LIC taxonomy's semantic content from mere label co-occurrence. That said, the paper's actual argument is about retrieval performance, not about the perceptual reality of LIC classes, and the body already concedes the patent-level cost. So the stress-test lands mostly on the abstract, not on the method's internal logic.

Who is this for? Anyone working on patent image retrieval or on hierarchical contrastive losses with cheap taxonomy labels. It's a modest incremental contribution, not a breakthrough, but the formulation is clean and the experimental breadth is useful. I'd send it to a serious referee. The fixes are easy: correct the abstract, add a sensitivity analysis for the weights, and ideally a shuffled-label control. Without those, the result is still a reasonable baseline, just with an overstated headline.

Referee Report

3 major / 3 minor

Summary. The paper proposes a hierarchical multi-positive contrastive learning (HMCL) loss for patent image retrieval. Instead of treating only same-patent images as positives, Eq. (4) weights every pair in the batch by a score h_ij from Eq. (2), which assigns decreasing relevance to images sharing the same patent ID, subclass, or main class under the Locarno International Classification (LIC). The authors fine-tune ViT, ResNet, and CLIP models on DeepPatent2 and report mAP/nDCG at three hierarchical levels, concluding that the method improves retrieval, especially for low-parameter models.

Significance. The loss formulation is simple, coherent, and a natural extension of standard contrastive learning, and the paper's strengths include multi-seed experiments across several architectures, use of the public DeepPatent2 benchmark, and a clear hierarchical evaluation protocol. If the claims were fully supported, the method would be a useful, low-cost contribution to patent image retrieval. However, the evidence is currently mixed: patent-ID mAP drops on several models, and the mechanism is not isolated from label co-occurrence, so the central claim is not yet established as stated.

major comments (3)

[Table 1 and Table 2] The abstract and Section 5 claim that HMCL 'enhances retrieval performance at all hierarchical levels,' but the patent-ID mAP in Table 1 decreases for ViT-Small (0.349 to 0.347), ViT-Base (0.333 to 0.324), and ViT-Large (0.348 to 0.345) relative to CL, and in Table 2 for CLIP-B/16* (0.373 to 0.356), CLIP-L/14* (0.454 to 0.452), CLIP-B/16 (0.401 to 0.386), and CLIP-L/14 (0.458 to 0.439). This directly contradicts the unqualified claim that the method enhances retrieval results. The paper must either restrict the claim to the subclass and main-class levels or characterize the trade-off explicitly, and the explanation in Section 4 that this deterioration is 'expected' should be supported by a mechanism check rather than asserted.
[Section 3, Eq. (2) and Eq. (5)] The scores sp=1, ss=0.35, sm=0.2 and the weight λ=0.2 are fixed without any sensitivity analysis, although the entire method rests on these hand-set hyperparameters. Please provide an ablation over these values (at least over ss and sm, and ideally over λ and the temperature τ) to show that the reported subclass/main-class gains are not an artifact of a particular weighting. Without such an ablation, it is unclear whether the method is robust or tuned to one setting.
[Section 3 and Section 4] The LIC taxonomy is used both to define the training positives (Eq. 2) and to define the evaluation relevance sets, so the observed subclass/main-class improvements could reflect fitting to administrative label structure rather than visual-semantic similarity. The paper needs a control experiment that breaks the label-signal link—for example, randomly permuting subclass and main-class labels during training, or training with non-hierarchical co-class positives—to establish that the gains come from the LIC hierarchy and not from label co-occurrence alone. This is load-bearing because the entire motivation of the method is the semantic usefulness of the taxonomy.

minor comments (3)

[Keywords and Section 5] The keyword 'Hierarchical Multipositive Contrastive Learning' and the phrase 'hierarchical multipositive contrastive learning' in Section 5 use 'Multipositive' as one word, while the abstract and title use 'multi-positive.' Please use hyphenated spelling consistently throughout.
[Eq. (4) and Eq. (5)] State explicitly whether the sums over j include the anchor itself. The standard contrastive loss excludes the self-pair, and the notation should be unambiguous about this; otherwise the normalization factor H_i in Eq. (3) cannot be verified.
[Section 4] The paragraph reporting standard deviations gives approximate aggregate values but does not report them per model or per method. Please include per-method standard deviations in the tables or a supplementary table, since the paper claims multi-seed experiments and the reader cannot assess the significance of the differences.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HMCL loss is a new objective evaluated on an external benchmark; results are empirical rather than derived from fitted constants.

full rationale

The paper's derivation chain is self-contained and does not reduce any claimed result to its own inputs. Equation 2 defines the hierarchical relevance scores h_ij using the LIC taxonomy, and Equation 4 inserts these scores into a multi-positive contrastive loss. The scores sp=1, ss=0.35, sm=0.2 are hand-set hyperparameters, not fitted to the test labels or to the retrieval metrics. The central claim that HMCL improves retrieval is supported by comparisons against a conventional contrastive baseline (CL) and a no-fine-tuning baseline on the external DeepPatent2 dataset. The authors explicitly report that patent-ID mAP declines for some models while subclass and main-class mAP improve, describing this as a trade-off rather than hiding it. Using the same taxonomy for training and evaluation is a standard supervised-learning setup and could raise a correctness question about whether gains reflect administrative labels or visual semantics, but that is not a circular derivation: no equation in the paper is equivalent to its inputs by construction, no fitted parameter is renamed as a prediction, and the cited works are used for context or baselines rather than as the sole justification of the central claim. The claim that Equation 1 is a special case of Equation 4 is a mathematical containment, not circularity. Overall, no circular step was found.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central training objective depends on four hand-set scalars (sp, ss, sm, lambda) and the assumption that LIC co-class images are useful positives. No new entities are postulated. These parameters are not fitted to test labels, but they are not justified by an ablation either.

free parameters (5)

sp = 1
Relevance score for same patent ID in Eq. 2; hand-set without sensitivity analysis.
ss = 0.35
Relevance score for same subclass in Eq. 2; hand-set and not ablated.
sm = 0.2
Relevance score for same main class in Eq. 2; hand-set and not ablated.
lambda = 0.2
Weight for language supervision term in Eq. 5; fixed without sensitivity study.
tau = 0.1
Temperature for contrastive loss; standard hyperparameter, but fixed and not studied.

assumptions (4)

domain assumption Cosine similarity in the embedding space is an appropriate relevance measure for patent images at all hierarchical levels.
Equations 1 and 4 use cosine similarity, but no analysis shows it matches human relevance judgments for technical drawings.
domain assumption Coarser LIC levels (same main class, same subclass) imply relevant positive relationships, with a fixed transitive ordering s_p > s_s > s_m.
Eq. 2 asserts relevance decreases with taxonomic distance, but no empirical validation shows that main-class matches are useful positives.
domain assumption The DeepPatent2 2007 subset is representative of patent image retrieval, and the given split yields unbiased evaluation.
Section 3 uses only 2007 data; generalizability to other years and classes is not tested.
standard math Standard softmax-normalized InfoNCE-style loss is a valid training objective for retrieval.
The paper relies on the standard contrastive learning framework of Eq. 1 without re-deriving it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hierarchical Multi-Positive Contrastive Learning for Patent Image Retrieval." pith.science (2026). https://pith.science/paper/G4TKDPWJ

@misc{pith2026250613496,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Multi-Positive Contrastive Learning for Patent Image Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G4TKDPWJ}},
  note         = {Machine review of arXiv:2506.13496}
}

read the original abstract

Patent images are technical drawings that convey information about a patent's innovation. Patent image retrieval systems aim to search in vast collections and retrieve the most relevant images. Despite recent advances in information retrieval, patent images still pose significant challenges due to their technical intricacies and complex semantic information, requiring efficient fine-tuning for domain adaptation. Current methods neglect patents' hierarchical relationships, such as those defined by the Locarno International Classification (LIC) system, which groups broad categories (e.g., "furnishing") into subclasses (e.g., "seats" and "beds") and further into specific patent designs. In this work, we introduce a hierarchical multi-positive contrastive loss that leverages the LIC's taxonomy to induce such relations in the retrieval process. Our approach assigns multiple positive pairs to each patent image within a batch, with varying similarity scores based on the hierarchical taxonomy. Our experimental analysis with various vision and multimodal models on the DeepPatent2 dataset shows that the proposed method enhances the retrieval results. Notably, our method is effective with low-parameter models, which require fewer computational resources and can be deployed on environments with limited hardware.

Figures

Figures reproduced from arXiv: 2506.13496 by the authors.

**Figure 1.** Overview of the Hierarchical Multi-Positive Contrastive Learning approach using the Locarno International Classification system. Figure (a) is an example of the hierarchical taxonomy of patents provided by the Locarno International Classification system. The images from Patent 2 and Patent 3 are omitted. Figure (b) presents how positive pairs within a batch are defined in conventional contrastive learning approaches… view at source ↗

**Figure 2.** reports the results with ResNet-18 and ResNet50 using the metrics MRR@K and Acc@K for 𝐾 ∈ {1, 5, 10, 20}, providing a more comprehensive overview of the retrieved list. For all levels (Patent ID, Subclass, and Main Class), the proposed approach outperforms the conventional contrastive learning method, with more relevant items being found at higher ranks in the retrieved list. Finally, we project the embeddings of … view at source ↗

**Figure 3.** First and second principal component of the embeddings of 5 different subclasses from the LIC system for the conventional Contrastive Learning (CL) and the Hierarchical MultiPositive Contrastive Learning (HMCL) method. The Subclass 1402 and 1403 from LIC belong to the same Main Class, while the rest to different ones. The red boxes enclose the main regions of Subclass 1402 and 1403. are inherently more suitable f… view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 20 canonical work pages

[1]

Kucer, D

M. Kucer, D. Oyen, J. Castorena, J. Wu, Deeppatent: Large scale patent drawing recognition and retrieval, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022, pp. 2309– 2318

work page 2022
[2]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Atten- tion is all you need, Advances in neural information processing systems 30 (2017)

work page 2017
[3]

Devlin, M.-W

J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for lan- guage understanding, in: Proceedings of the 2019 conference of the North American chapter of the asso- 4 Kshitij Kavimandan et al.CEUR Workshop Proceedings1–6 ciation for computational linguistics: human language technologies, volume 1 (long and sh...

work page 2019
[4]

Dosovitskiy, L

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weis- senborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Min- derer, G. Heigold, S. Gelly, J. Uszkoreit, N. Houlsby, An image is worth 16x16 words: Transformers for image recognition at scale, in: International Confer- ence on Learning Representations, 2021. URL: https: //openreview.net/forum?id=YicbFdNTTy

work page 2021
[5]

K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE con- ference on computer vision and pattern recognition, 2016, pp. 770–778

2016
[6]

Geirhos, P

R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, W. Brendel, Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, in: International conference on learning representations, 2018

work page 2018
[7]

H. Wang, S. Ge, Z. Lipton, E. P. Xing, Learning robust global representations by penalizing local predictive power, Advances in neural information processing systems 32 (2019)

work page 2019
[8]

Sangkloy, N

P. Sangkloy, N. Burnell, C. Ham, J. Hays, The sketchy database: learning to retrieve badly drawn bunnies, ACM Transactions on Graphics (TOG) 35 (2016) 1–12

work page 2016

Show all 26 references

[9]

Aubakirova, K

D. Aubakirova, K. Gerdes, L. Liu, Patfig: Generating short and long captions for patent figures, in: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 2843–2849

2023
[10]

Lo, J.-M

H.-C. Lo, J.-M. Chu, J. Hsiang, C.-C. Cho, Large lan- guage model informed patent image retrieval,
[11]

Radford, J

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

2021
[12]

Ajayi, X

K. Ajayi, X. Wei, M. Gryder, W. Shields, J. Wu, S. M. Jones, M. Kucer, D. Oyen, Deeppatent2: A large-scale benchmarking corpus for technical drawing under- standing, Scientific Data 10 (2023) 772

2023
[13]

H. Wang, Y. Zhang, Learning efficient representations for image-based patent retrieval, in: Chinese Con- ference on Pattern Recognition and Computer Vision (PRCV), Springer, 2023, pp. 15–26

2023
[14]

Higuchi, K

K. Higuchi, K. Yanai, Patent image retrieval using transformer-based deep metric learning, World Patent Information 74 (2023) 102217

2023
[15]

World Intellectual Property Office, Locarno classifi- cation, https://www.wipo.int/classifications/locarno/,
[16]

Mettes, M

P. Mettes, M. Ghadimi Atigh, M. Keller-Ressel, J. Gu, S. Yeung, Hyperbolic deep learning in computer vision: A survey, International Journal of Computer Vision 132 (2024) 3484–3508

2024
[17]

Nalmpantis, P

A. Nalmpantis, P. Lippe, S. Magliacane, Hierarchical causal representation learning, in: Causal Representa- tion Learning Workshop at NeurIPS 2023, 2023

2023
[18]

Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, B. Guo, Swin transformer: Hierarchical vision trans- former using shifted windows, in: Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10012–10022

2021
[19]

Y. Tian, L. Fan, P. Isola, H. Chang, D. Krishnan, Sta- blerep: Synthetic images from text-to-image models make strong visual representation learners, Advances in Neural Information Processing Systems 36 (2023) 48382–48402

2023
[20]

van den Oord, Y

A. van den Oord, Y. Li, O. Vinyals, Represen- tation learning with contrastive predictive cod- ing, 2019. URL: https://arxiv.org/abs/1807.03748. arXiv:1807.03748

2019 arXiv
[21]

Loshchilov, F

I. Loshchilov, F. Hutter, Decoupled weight decay regu- larization, in: International Conference on Learning Representations, 2019. URL: https://openreview.net/ forum?id=Bkg6RiCqY7

2019
[22]

Paszke, S

A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Rai- son, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, Pytorch: An imperative style, high- perf...

2019 arXiv
[23]

Falcon, The PyTorch Lightning team, Py- Torch Lightning, 2019

W. Falcon, The PyTorch Lightning team, Py- Torch Lightning, 2019. URL: https://github.com/ Lightning-AI/lightning. doi: 10.5281/zenodo. 3828935

2019 doi
[24]

T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y. Jernite, J. Plu, C. Xu, T. Le Scao, S. Gugger, M. Drame, Q. Lhoest, A. Rush, Transformers: State-of-the-art natural lan- ...

2020
[2024]

arXiv:2404.19360

URL: https://arxiv.org/abs/2404.19360. arXiv:2404.19360

arXiv
[2025]

Accessed: April 2025

2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Kucer, D

M. Kucer, D. Oyen, J. Castorena, J. Wu, Deeppatent: Large scale patent drawing recognition and retrieval, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022, pp. 2309– 2318

work page 2022

[2] [2]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Atten- tion is all you need, Advances in neural information processing systems 30 (2017)

work page 2017

[3] [3]

Devlin, M.-W

J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for lan- guage understanding, in: Proceedings of the 2019 conference of the North American chapter of the asso- 4 Kshitij Kavimandan et al.CEUR Workshop Proceedings1–6 ciation for computational linguistics: human language technologies, volume 1 (long and sh...

work page 2019

[4] [4]

Dosovitskiy, L

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weis- senborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Min- derer, G. Heigold, S. Gelly, J. Uszkoreit, N. Houlsby, An image is worth 16x16 words: Transformers for image recognition at scale, in: International Confer- ence on Learning Representations, 2021. URL: https: //openreview.net/forum?id=YicbFdNTTy

work page 2021

[5] [5]

K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE con- ference on computer vision and pattern recognition, 2016, pp. 770–778

2016

[6] [6]

Geirhos, P

R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, W. Brendel, Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, in: International conference on learning representations, 2018

work page 2018

[7] [7]

H. Wang, S. Ge, Z. Lipton, E. P. Xing, Learning robust global representations by penalizing local predictive power, Advances in neural information processing systems 32 (2019)

work page 2019

[8] [8]

Sangkloy, N

P. Sangkloy, N. Burnell, C. Ham, J. Hays, The sketchy database: learning to retrieve badly drawn bunnies, ACM Transactions on Graphics (TOG) 35 (2016) 1–12

work page 2016