ContextRefine-CLIP for EPIC-KITCHENS-100 Multi-Instance Retrieval Challenge 2025

Jing He; Kexin Zhang; Lingling Li; Puhua Chen; Yiqing Wang

REVIEW 4 major objections 5 minor 12 references

ContextRefine-CLIP for EPIC-KITCHENS-100 Multi-Instance Retrieval Challenge 2025

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Bidirectional cross-modal attention before projection lifts EPIC-KITCHENS-100 retrieval to 66.8 mAP and 82.1 nDCG in a single model.

desk verdict Reasonable incremental architecture with a believable ablation, but the causal claim needs a proper baseline control and a fix to Eq. (5) before I'd trust the +4.1 mAP. read the letter →

arxiv 2506.10550 v1 pith:GYPZL2QA submitted 2025-06-12 cs.CV

Jing He , Yiqing Wang , Lingling Li , Kexin Zhang , Puhua Chen This is my paper · ORCID

classification cs.CV

keywords multi-instanceretrievalcross-modalEPIC-KITCHENS-100video-textcross-attentionsoftlabelsSymmetricMulti-SimilarityLosstest-timeaugmentation

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 establish that a relatively small architectural change to a CLIP-style dual encoder, inserting a bidirectional cross-modal attention module between the visual and text encoders and the projection heads, materially improves multi-instance retrieval when relevance comes as a soft-label matrix rather than hard binary matches. On EPIC-KITCHENS-100, the resulting ContextRefine-CLIP (CR-CLIP) raises average mAP from 62.1% to 66.8% and average nDCG from 73.0% to 82.1% over the SMSLoss baseline, with the CMCR module alone contributing 4.1 mAP and 7.9 nDCG before any test-time augmentation. Combined with horizontal-flip and multi-scale test-time augmentation, the single model reaches 66.78 mAP and 82.08 nDCG on the challenge's public leaderboard without ensemble learning. A reader should care because the result suggests that feature-level inter-modal interaction, not just a better loss function, is what drives ranking quality in soft-label retrieval.

What carries the argument

The load-bearing component is the Cross-Modal Context Refinement (CMCR) module: a pair of lightweight cross-attention layers, one that lets text features attend to visual features and one that does the symmetric operation for vision-guided text refinement. Each refined feature passes through a residual connection, LayerNorm, and a Gated Feed-Forward Network before the MLP projection heads. The module's job is to make each modality's representation absorb semantic cues from the other before similarity is computed, so that fine-grained correspondences such as video actions to query verbs and object combinations to descriptions are aligned in the shared embedding space. The paper argues this is what produces the gains over a projection-only dual encoder, and it is trained with Symmetric Multi-Similarity Loss to exploit the soft-label relevance matrix.

What would settle it

Retrain the SMSLoss baseline from the same AVION/Ego4D pretrained checkpoint, with the same batch size (160), learning rate (1.8e-5), ViT-L-14 backbone, frame sampling, and flip/multi-scale TTA protocol, and compare on the same EPIC-KITCHENS-100 splits; if the mAP/nDCG gap is not approximately 4.1/7.9, the causal attribution to the CMCR module fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that on EPIC-KITCHENS-100 multi-instance retrieval, a dual-encoder model whose visual and textual features refine each other through cross-attention before being projected into a shared space outperforms the prior SMSLoss approach by a wide margin. In the paper's comparison table, CR-CLIP with the CMCR module reaches 66.2% average mAP and 80.9% average nDCG, surpassing the SMSLoss baseline (62.1% mAP, 73.0% nDCG), and adding GatedFFN, horizontal flip, and multi-scale test-time augmentation brings it to 66.8% average mAP and 82.1% average nDCG. The authors further report that the same single model, without ensembling, scores 66.78% mAP and 82.08% nDCG on the public leaderboard, beating the 2024 ensemble-based solution's 63.8% mAP and 74.3% nDCG.

Load-bearing premise

The load-bearing premise is that the Table 1 SMSLoss baseline was trained with the same pretrained initialization, data sampling, frame selection, and evaluation protocol as CR-CLIP, so the measured 4.1 mAP and 7.9 nDCG differences can be attributed to the CMCR module; the paper does not specify this.

Editorial extensions

If this is right

A single CR-CLIP model without ensembling reaches 66.78 mAP and 82.08 nDCG on the EPIC-KITCHENS-100 public leaderboard, above the 2024 ensemble baseline's 63.8 mAP and 74.3 nDCG.
The CMCR module alone adds 4.1 average mAP and 7.9 average nDCG over the SMSLoss baseline, so the architectural change, not just the loss or augmentation, drives most of the gain.
GatedFFN improves T2V mAP and average nDCG, and flip plus multi-scale TTA adds a further 0.6 average mAP and 1.2 average nDCG on top of CMCR.
The model trains on a single machine with 4 RTX 3090s and batch size 160, so the approach is reproducible without large compute clusters.
Because the largest relative improvement is in nDCG, the refined features mainly improve the ordering of retrieved items, not just hit-or-miss accuracy.

Reading between the lines

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

The same bidirectional refinement idea could transfer to other soft-label retrieval tasks such as composed image retrieval or video moment retrieval, but the paper only tests it on EPIC-KITCHENS-100.
The paper's ablations combine CMCR, GatedFFN, and TTA in a chain, so the individual contribution of multi-scale augmentation versus the module itself on the final leaderboard number is not isolated; a factorial ablation would settle it.
The headline comparison in the introduction is against a 2024 ensemble solution (63.8/74.3), while Table 1's baseline is SMSLoss (62.1/73.0); readers should compare like-for-like when quoting the +3.0/+7.8 improvement.
The authors' own limitation list notes no ensembling, no computational-overhead analysis, ViT-L only, and limited hyperparameter tuning, so the reported numbers are a single configuration rather than an optimized family.

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 solid competition report with a sensible incremental idea, but the headline '+4.1 mAP from CMCR' is not actually pinned down by the experiments as written. The ablation trend is plausible and the leaderboard result is real, but the missing protocol details for the SMSLoss baseline and a dimension mismatch in Eq. (5) keep the central attribution from being fully supported.

What's actually new: inserting a lightweight cross-attention module (CMCR) between the encoders and the projection heads of AVION, plus GatedFFN and multi-scale TTA, and showing consistent increments on EK-100 multi-instance retrieval. That's a reasonable extension of known parts. The ablation table is monotonic (SMSLoss -> +CMCR -> +GatedFFN -> +Flip -> +Scale), which is a good sign. The paper is honest about its limitations: no ensembling, no overhead analysis, only ViT-L. That counts for something. The code is promised, not yet available.

Soft spots: the biggest is the baseline comparability. Table 1's SMSLoss row is the comparison that matters, but the text never says whether that row was produced by re-running the SMSLoss protocol under identical initialization, sampling, frames, and evaluation, or copied from the prior paper. The intro instead compares against a 2024 ensemble baseline, which is a different target. Without that control, the +4.1 mAP could be partly pipeline differences. That's a real gap, though fixable by a careful re-run or a clear statement. Second, Eq. (5) adds F_t to the cross-attention output which lives in attention dimension D_a; the residual only makes sense if D_t = D_a, which is never stated. That is a specification bug that matters for reproducibility. Minor: no error bars or significance tests; typical for challenge reports, so I wouldn't weight it heavily.

The stress-test note is on target. I agree with the reader's conditional verdict. The central idea is defensible; the evidence is just incomplete, not contradictory.

Bottom line: if you care about EK-100 retrieval or CLIP-style video-text dual encoders, this is a useful data point. It deserves a serious referee (workshop or short paper) because the result is plausible and the ablations are informative, but it should be sent back for a baseline-provenance statement, a corrected Eq. (5), and ideally released code before acceptance.

Referee Report

4 major / 5 minor

Summary. This paper presents ContextRefine-CLIP (CR-CLIP), a dual-encoder model designed for the EPIC-KITCHENS-100 Multi-Instance Retrieval Challenge 2025. The method extends the AVION architecture with a Cross-Modal Context Refinement (CMCR) module that performs bidirectional cross-attention between visual and textual features, followed by a Gated Feed-Forward Network (GatedFFN) and LayerNorm residual blocks. Training uses the Symmetric Multi-Similarity Loss on a pretrained checkpoint built from SMSLoss and AVION. At inference, features are aggregated under horizontal flipping and multi-scale resizing (0.875x, 1.0x, 1.125x). The authors report a single-model, non-ensemble result of 66.78 mAP and 82.08 nDCG on the public leaderboard, stating that it achieves 1st place. Table 1 shows an ablation from SMSLoss to CR-CLIP (+4.1 mAP, +7.9 nDCG) and further small increments from GatedFFN, flip, and multi-scale test-time augmentation.

Significance. If the reported leaderboard numbers hold, this is a strong empirical result for the EK-100 MIR benchmark and demonstrates that cross-modal attention refinement before projection can improve a CLIP-style dual encoder for multi-instance retrieval with soft labels. The paper provides a clear ablation chain, reports an external leaderboard result, and promises code release, which are concrete strengths. However, the scientific contribution is primarily empirical and incremental; there is no theoretical analysis, and the novelty relative to existing cross-attention fusion designs is limited. The main value is in validating the CMCR module and the enhanced TTA strategy on a challenging benchmark, provided the baseline comparison is like-for-like.

major comments (4)

[Section 3.2, Table 1 vs. Introduction] The causal attribution of the reported gains to the CMCR module requires that the SMSLoss row in Table 1 is a like-for-like baseline produced under the identical protocol as CR-CLIP (same pretrained initialization, same frame sampling, same batch composition, same training length, same evaluation). Section 3.1 only says 'the same training setup as SMSLoss' and does not state whether this baseline was re-run by the authors or reused from the prior SMSLoss paper. This ambiguity is compounded by the Introduction comparing CR-CLIP to a different '2024 ensemble' baseline (63.8 mAP, 74.3 nDCG) rather than the Table 1 SMSLoss row (62.1 mAP, 73.0 nDCG). Please specify exactly which baseline is the control for each claim and report the protocol details (epochs, frames, resolution, batch composition) for both the baseline and the proposed method.
[Section 2.2, Eq. (5)] As written, Eq. (5) is dimensionally inconsistent: F_t ∈ R^{B×Dt} is added to F_{v←t}, but F_{v←t} is the output of the multi-head attention with query Q_t ∈ R^{B×Da} and keys/values V_v ∈ R^{B×Da}, so the residual requires D_t = D_a, which is never stated. If the intended residual is on the projected query Q_t, or if the architecture sets D_t = D_a by construction, please rewrite the notation to make the dimensions explicit. As it stands, the module is not precisely specified enough to reproduce from the equations.
[Section 3.2, Table 1] The per-component ablation increments are small (e.g., +0.2 mAP from 'w/Flip' to 'w/Flip+Scale' and a drop of 0.6 mAP in V2T from 'CR-CLIP' to 'w/GatedFFN'), yet the text claims these components play 'a crucial role.' Without error bars, multiple seeds, or significance tests, these small differences cannot be distinguished from training noise. Please report variance or at least state that these increments are single-run observations.
[Section 2.2, Eq. (6)] The GatedFFN is not defined: Eq. (6) refers to GatedFFN(F'_v) but provides no equation or description of the gating mechanism, the dimensionality, the activation functions, or how it differs from a standard FFN. The symmetric vision-guided text refinement operation is also not shown. This prevents reproduction of the method independent of the promised code release.

minor comments (5)

[Section 3] The heading 'Experimemts' contains a typo and should read 'Experiments.'
[Abstract vs. Table 1] The abstract reports 66.78 mAP and 82.08 nDCG, while Table 1 reports 66.8 and 82.1; the phrase 'retained the first three valid digits without rounding' is ambiguous and should be clarified (e.g., significant figures vs. decimal places).
[Section 2.4] The multi-scale augmentation '0.875x, 1.0x and 1.125x of the original resolution' should specify the reference resolution for the base case and how center cropping interacts with the original aspect ratio.
[References] Reference [9] appears unrelated to the sentence citing it on contrastive representation learning; please verify the citation placement and relevance.
[Section 2.3] Eq. (7) defines only the difference R = C_ij - C_ik and does not provide the SMSLoss formula or define the role of the relaxation factor τ; add the full loss expression or a precise citation to make the objective reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: results are external leaderboard measurements with an architectural ablation, not a derivation that reduces to its inputs.

full rationale

The paper's central claim is an empirical performance measurement on the EPIC-KITCHENS-100 public leaderboard, with ablations over the proposed CMCR module, GatedFFN, and test-time augmentation. The gains are attributed to an architectural module trained with an external loss (SMSLoss) and an external pretrained initialization (AVION), not to parameters fitted to the target metric or to a self-citation chain. The CMCR module is defined as standard cross-attention with residual and gated FFN layers; the reported numbers are measured outputs, not consequences of the module's definition. The only suspicious passages are an experimental-protocol ambiguity (Table 1's SMSLoss row is not shown to be produced under the identical pipeline, despite the text claiming 'the same training setup as SMSLoss') and a dimension inconsistency in Eq. (5), where F_t is added to F_{v←t}; these are correctness/reproducibility concerns, not circularity, because they do not make the empirical result equivalent to its input by construction. The paper also explicitly lists limitations (no ensembling, no overhead analysis, ViT-L only, limited hyperparameter search), which further supports that no hidden self-referential justification is being used. No load-bearing self-citation or uniqueness-imported-from-authors pattern is present.

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

The paper introduces no new physical entities or conceptual constructs; it is an empirical engineering report. The central claim rests on the transferability of the AVION pretrained model, the reliability of the provided soft labels, the correctness of the described (but not fully specified) CMCR implementation, and the fairness of the comparison to the SMSLoss baseline.

free parameters (4)

learning rate = 1.8e-5
Manually set to match the SMSLoss setup; no sensitivity analysis is reported, and the final result depends on this choice.
batch size = 160
Chosen to fit 4x RTX 3090 GPUs with 40 clips per GPU; no ablation on batch size is given.
CMCR hyperparameters = attention dim 512, 8 heads, dropout 0.1
Hand-selected architecture dimensions; no tuning study is reported.
TTA scales = 0.875x, 1.0x, 1.125x
Hand-chosen multi-scale augmentation factors; the choice is not justified experimentally.

assumptions (4)

domain assumption The pretrained AVION model, trained on Ego4D with LLM-augmented captions, transfers to EPIC-KITCHENS-100 and provides a suitable initialization such that the effects observed are attributable to the new module.
Section 3.1 builds directly on the AVION pretrained model. If the pretrained representation does not transfer, the reported gains cannot be attributed to CMCR.
domain assumption The soft-label relevance matrix C in EPIC-KITCHENS-100 is a reliable supervision signal and the Symmetric Multi-Similarity Loss, through R = C_ij - C_ik, correctly optimizes the model for this soft-label setting.
Section 2.3 relies entirely on SMSLoss and the provided soft labels. If the labels are noisy or the loss formulation is inappropriate, the training objective would be invalid.
ad hoc to paper The residual connection in the cross-attention module is implemented as intended despite the dimension mismatch in Eq. (5).
Eq. (5) writes F'_v = LayerNorm(F_t + F_{v<-t}), adding a text feature to a visual refinement output. Unless additional projections are omitted from the text, the module described cannot be implemented as written.
domain assumption The evaluation protocol for the public leaderboard is identical for CR-CLIP and all compared baselines, including the same data splits, preprocessing, and metric computation.
Reported gains in Table 1 presume identical evaluation conditions. Any difference would break the comparison, especially since the introduction cites a different 2024 ensemble baseline.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ContextRefine-CLIP for EPIC-KITCHENS-100 Multi-Instance Retrieval Challenge 2025." pith.science (2026). https://pith.science/paper/GYPZL2QA

@misc{pith2026250610550,
  author       = {Pith},
  title        = {Pith review of: ContextRefine-CLIP for EPIC-KITCHENS-100 Multi-Instance Retrieval Challenge 2025},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GYPZL2QA}},
  note         = {Machine review of arXiv:2506.10550}
}

read the original abstract

This report presents ContextRefine-CLIP (CR-CLIP), an efficient model for visual-textual multi-instance retrieval tasks. The approach is based on the dual-encoder AVION, on which we introduce a cross-modal attention flow module to achieve bidirectional dynamic interaction and refinement between visual and textual features to generate more context-aware joint representations. For soft-label relevance matrices provided in tasks such as EPIC-KITCHENS-100, CR-CLIP can work with Symmetric Multi-Similarity Loss to achieve more accurate semantic alignment and optimization using the refined features. Without using ensemble learning, the CR-CLIP model achieves 66.78mAP and 82.08nDCG on the EPIC-KITCHENS-100 public leaderboard, which significantly outperforms the baseline model and fully validates its effectiveness in cross-modal retrieval. The code will be released open-source on https://github.com/delCayr/ContextRefine-Clip

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 9 canonical work pages

[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]

Scaling egocentric vision: The epic-kitchens dataset

Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European conference on computer vision (ECCV), pages 720--736, 2018

work page 2018
[3]

Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100

Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, pages 1--23, 2022

work page 2022
[4]

Ego4d: Around the world in 3,000 hours of egocentric video

Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995--19012, 2022

2022
[5]

A survey on vision transformer

Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE transactions on pattern analysis and machine intelligence, 45 0 (1): 0 87--110, 2022

work page 2022
[6]

Contrastive representation learning: A framework and review

Phuc H Le-Khac, Graham Healy, and Alan F Smeaton. Contrastive representation learning: A framework and review. Ieee Access, 8: 0 193907--193934, 2020

work page 2020
[7]

Visual to text: Survey of image and video captioning

Sheng Li, Zhiqiang Tao, Kang Li, and Yun Fu. Visual to text: Survey of image and video captioning. IEEE Transactions on Emerging Topics in Computational Intelligence, 3 0 (4): 0 297--312, 2019

work page 2019
[8]

Learning transferable visual models from natural language supervision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748--8763. PmLR, 2021

2021

Show all 12 references

[9]

Symmetric multi-similarity loss for epic-kitchens-100 multi-instance retrieval challenge 2024, 2024 a

Xiaoqi Wang, Yi Wang, and Lap-Pui Chau. Symmetric multi-similarity loss for epic-kitchens-100 multi-instance retrieval challenge 2024, 2024 a

2024
[10]

Variance alignment score: A simple but tough-to-beat data selection method for multimodal contrastive learning

Yiping Wang, Yifang Chen, Wendan Yan, Kevin Jamieson, and Simon Shaolei Du. Variance alignment score: A simple but tough-to-beat data selection method for multimodal contrastive learning. arXiv preprint arXiv:2402.02055, 2024 b

2024 arXiv
[11]

a henb \

Yue Zhao and Philipp Kr \"a henb \"u hl. Training a large video model on a single machine in a day, 2023

2023
[12]

a henb \

Yue Zhao, Ishan Misra, Philipp Kr \"a henb \"u hl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586--6597, 2023

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[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]

Scaling egocentric vision: The epic-kitchens dataset

Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European conference on computer vision (ECCV), pages 720--736, 2018

work page 2018

[3] [3]

Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100

Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision, pages 1--23, 2022

work page 2022

[4] [4]

Ego4d: Around the world in 3,000 hours of egocentric video

Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995--19012, 2022

2022

[5] [5]

A survey on vision transformer

Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE transactions on pattern analysis and machine intelligence, 45 0 (1): 0 87--110, 2022

work page 2022

[6] [6]

Contrastive representation learning: A framework and review

Phuc H Le-Khac, Graham Healy, and Alan F Smeaton. Contrastive representation learning: A framework and review. Ieee Access, 8: 0 193907--193934, 2020

work page 2020

[7] [7]

Visual to text: Survey of image and video captioning

Sheng Li, Zhiqiang Tao, Kang Li, and Yun Fu. Visual to text: Survey of image and video captioning. IEEE Transactions on Emerging Topics in Computational Intelligence, 3 0 (4): 0 297--312, 2019

work page 2019

[8] [8]

Learning transferable visual models from natural language supervision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748--8763. PmLR, 2021

2021