Pith. sign in

REVIEW 4 major objections 5 minor 29 references

A Few-Shot Metric Learning Method with Dual-Channel Attention for Cross-Modal Same-Neuron Identification

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a few-shot metric learning network with dual-channel attention identifies the same neuron across two-photon and fMOST images, reaching 77.4% recall and 90.1% specificity on held-out pairs.

desk verdict A plausible engineering prototype for cross-modal neuron matching, but the headline recall/specificity numbers rest on a test set partly built from the model's own similarity scores and need a rework before they can be trusted. read the letter →

arxiv 2504.16520 v1 pith:TM44R4CZ submitted 2025-04-23 cs.CV q-bio.NC

Wenwei Li , Liyi Cai , Wu Chen , Anan Li This is my paper · ORCID
classification cs.CVq-bio.NC
keywords neuronsimilaritylearningmetriccross-modalmatchingdual-channelattentionhardsampleminingCircleLosstwo-photonimagingfMOST
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that identifying the same neuron across two brain-imaging modalities—two-photon functional imaging and fMOST, a fluorescence sectioning technique that reveals whole-brain structure—can be automated with very few annotated examples. It proposes a few-shot metric learning method that maps image patches from both modalities into one shared embedding space, starting from a pretrained vision transformer and fine-tuning on 190 manually matched neuron pairs. On 53 held-out matched pairs against 212 unmatched pairs, the model reports 77.4% recall while rejecting 90.1% of unmatched pairs, and its Top-K retrieval curve beats conventional similarity metrics, single-channel ablations, and several fine-tuning baselines. A careful reader would care because single-neuron matching between functional and structural images is currently a manual bottleneck for linking what a neuron does with how it is wired.

What carries the argument

The load-bearing object is the dual-channel attention network: one channel applies attention to local soma morphology (contour shapes and fine textures), the other applies spatial attention over a pretrained Transformer's representation of fiber trajectories and surrounding structures, and a Gate module adaptively weights and sums the two streams into a single embedding. Around this sit the MultiSimilarityMiner, which during training dynamically selects negative pairs within a distance threshold $\alpha$ of an anchor, and Circle Loss, which weights gradients by how close samples sit to the decision margin using similarity scores $s_p^i$ and $s_n^j$ with margin $m$. Together they turn the task into a few-shot metric learning problem: the embedding itself is the identity model, and matching a novel neuron is nearest-neighbor search in that space.

What would settle it

Evaluate the trained model on a negative set drawn without any model-guided filtering—every unmatched two-photon/fMOST combination from a brain region, or random sampling only—and compare specificity at the same distance threshold; if the rejection rate falls well below 90.1%, the discriminative claim does not survive contact with the real unmatched-pair distribution. A complementary check reruns the same 53-query retrieval test after retraining with the hard-mining module disabled to see whether the reported Top-1 advantage persists.

Watch

Extended reading notes

Core claim

The central claim is that cross-modal neuron identity is learnable as a metric: after fine-tuning a pretrained vision transformer on only 190 matched two-photon/fMOST pairs, neurons from the two modalities can be embedded so that the same neuron lies close together and different neurons lie far apart. The network achieves this by splitting its attention into two channels—a local channel that reads soma contour and texture, and a global channel that reads fiber trajectories and surrounding context—and fusing the two streams with a learned gate. Hard-negative mining and Circle Loss push the embedding to concentrate on confusable pairs during training. Tested on 53 unseen matched pairs and 212 unmatched pairs, the model yields 41 true positives against 12 false negatives while rejecting 90.1% of negatives, and it retrieves the correct fMOST neuron at the top ranks more often than any compared baseline.

Load-bearing premise

The evaluation assumes that a negative test set built from 106 unmatched pairs the model itself finds hardest plus 106 random pairs represents the real population of unmatched neuron pairs; if real unmatched pairs differ in difficulty from this curated mix, the reported 90.1% rejection rate will not transfer to uncurated data.

Editorial extensions

If this is right

  • Single-neuron matching between functional and structural brain imaging becomes automatable from a few hundred annotated pairs, replacing manual registration in the two-photon/fMOST workflow.
  • Cross-modal retrieval becomes a practical query tool: a neuron imaged with two-photon microscopy can pull up its full fMOST morphology from a database, directly linking activity to structure.
  • Parameter-efficient fine-tuning of a general vision backbone is sufficient for neuron similarity, meaning small neuroscience datasets do not require training from scratch.
  • Both the soma-level channel and the fiber-context channel are needed: removing either branch measurably degrades Top-K retrieval, according to the ablations.
  • Hard negative mining and Circle Loss are what sharpen the decision boundary, since replacing either with random sampling or Triplet Loss lowers Top-K accuracy.

Reading between the lines

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

  • The local/global split is a transferable recipe: other cross-modal cell-matching problems may benefit from encoding a cell's intrinsic shape and its extrinsic context separately before fusion.
  • Because specificity is measured against a negative set deliberately enriched with the model's own near-misses, real-world rejection rates on uncurated pairs could move in either direction; an evaluation on randomly drawn pairs would settle which.
  • A learning-curve study—adding matched pairs in batches and watching recall per added pair—would show whether the few-shot advantage persists as annotations grow, which the paper does not test.
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

4 major / 5 minor

Summary. The paper proposes a few-shot metric learning framework for cross-modal neuron identification between two-photon and fMOST images. The method uses a pretrained vision transformer as the backbone, a dual-channel attention module (local soma morphology and global fiber context), a gated fusion mechanism, MultiSimilarityMiner-based hard negative mining, and Circle Loss. The authors report a recall of 77.4% and a specificity of 90.1% on a test set with 53 positive and 212 negative pairs, together with Top-k retrieval results, ablations, fine-tuning strategy comparisons, and attention/embedding visualizations.

Significance. The problem is timely and the architectural design is well motivated by the neuroscience setting: combining soma-level local detail with fiber-level global context is a sensible prior, and the few-shot pretrained-transfer approach is appropriate for the small annotated dataset. The paper includes useful ablations, Grad-CAM visualizations, t-SNE plots, and comparisons against several pixel-level similarity metrics. On the other hand, the experimental support for the headline claims is undermined by a test-set construction that uses the model's own predictions, an unspecified classification threshold, the absence of the most relevant automatic baseline from the literature, and the lack of any error bars or significance tests. If the evaluation is redone with a model-independent test protocol and appropriate baselines, the method could be a useful contribution; in its current form, the central quantitative claims are not established.

major comments (4)
  1. [Section 3.1, Table 1] The test negative set is constructed by selecting 106 'hard negatives' as the unmatched pairs with the highest predicted similarity scores according to the model being evaluated. This makes the negative test distribution a function of the model's own embedding, so the reported specificity (191/212 = 90.1%) is an estimate for an adversarially chosen distribution rather than for the real unmatched-pair distribution. The 50/50 mix of hard and random negatives is arbitrary, and the pooled specificity changes with that ratio. To support the claim, the authors should evaluate on a fixed negative set chosen independently of the model (e.g., a held-out random sample, or negatives defined before training) and report recall/specificity or ROC/AUC on that fixed set.
  2. [Section 3.1, Table 1] The classification threshold on the learned distance is never specified, and the paper reports only a single confusion matrix. It is therefore unclear whether the threshold was fixed on validation data or chosen post hoc on the test set. A single operating point without the threshold, ROC curve, or precision-recall curve does not substantiate the headline 77.4% recall and 90.1% specificity. The authors should report the threshold value, how it was selected, and full ROC/PR curves.
  3. [Section 3.6] The comparative evaluation is limited to pixel-level similarity metrics (NMI, cosine, SSIM, Pearson, MSE). The Introduction cites Li et al. [4], an automatic graph-based neuron-matching method for the same two-photon/fMOST task, but that method is not included as a baseline. Consequently, the claim that the proposed method outperforms 'existing methods' is not supported for the most relevant class of competing approaches. The authors should add the state-of-the-art automatic matching baseline from the cited literature and report quantitative comparisons.
  4. [Sections 3.3-3.4, Figures 4-5] All comparative claims ('significantly outperforms all ablation variants', 'consistently outperforms all baseline fine-tuning strategies') are based on Top-k accuracy curves without numerical values, error bars, independent runs, or significance tests. With a test set of only 53 positive pairs, differences could be within chance. The authors should report means and standard deviations over multiple seeds, or at least a paired statistical test, and provide the numerical results behind Figures 4 and 5.
minor comments (5)
  1. [Section 2.1] The paper does not state how unmatched pairs are generated or whether their constituent neurons are disjoint from the training pairs; this should be clarified to rule out data leakage in the negative set.
  2. [Section 2.3, Eq. (2)] Eq. (2) is not the standard Circle Loss formulation from [23] but a simplified pairwise softplus objective; the exact loss used should be clarified, and the values of alpha, m, gamma, and the LoRA ranks should be reported.
  3. [General] Figure 3 is missing; the text jumps from Figure 2 to Figure 4. The numbering of 'Table 3-8' is also unusual and should be corrected.
  4. [Section 3.4] There are grammatical errors and inconsistent capitalization in the text (e.g., 'it is evident that The original pretrained model fails to distinguish...'); the manuscript should be proofread.
  5. [Data and code availability] No code or data are provided; given the small dataset, releasing the annotated pairs and evaluation scripts would greatly improve reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

Evaluation loop: test negatives are selected using the model's own similarity scores, so the reported specificity and baseline comparisons are not independent of the model being validated.

  1. self definitional [Section 3.1 (Evaluation of Method Effectiveness), negative sample selection strategy, and the resulting Table 1 confusion matrix]
    "Hard negatives: 106 unmatched pairs are selected from non-matching neuron combinations that exhibit the highest predicted similarity scores (i.e., most likely to confuse the model), serving as challenging distractors. Random negatives: Another 106 unmatched pairs are randomly sampled from two-photon and fMOST neuron datasets, aimed at increasing diversity and generalizability of the evaluation set."

    The test negative distribution is constructed from the model's own predicted similarity ranking: 106 of 212 negatives are deliberately chosen as the unmatched pairs the model scores highest. The reported specificity (90.1% = 191/212) is therefore computed on a negative set that is a function of the very model being evaluated, rather than on a fixed or random sample of the real unmatched-pair distribution. The 50/50 hard/random mix is arbitrary, so the pooled specificity has no defined population target and changes with the mix ratio. When this same model-defined set is used in Section 3.6 to compare against NMI, cosine, SSIM, Pearson, and MSE baselines, the comparison is also biased: the hard negatives were selected to be hard for the proposed model's ranking, not for the baselines.

full rationale

The core methodological derivation is self-contained: Eq. (1) defines training-time hard negatives via the current feature embedding, and Eq. (2) is the standard Circle Loss; using a pretrained Unicom backbone is an external architectural choice, not a circular one. The self-citation to Li et al. [4] in the Introduction is a related-work contrast and is not load-bearing for the paper's claims. The central circularity is confined to the evaluation protocol in Section 3.1: the negative test set is defined by the model's own predicted similarity scores, so the headline numbers (77.4% recall, 90.1% specificity) and the Section 3.6 method comparisons are not independent validations of the model. This is not full circularity because the model is not fitting the test labels and the hard-negative selection makes the test conservative for the proposed model, but it makes the reported operating point and cross-method ranking depend on the model's own similarity ranking and on an arbitrary mix of hard and random negatives. The Discussion candidly acknowledges small dataset size and underexplored theory, but it does not address this test-selection loop, so the evaluation remains partially circular.

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

No new physical entities or scientific objects are postulated. The free parameters are unstated hyperparameters and thresholds, and the axioms are the domain-transfer and annotation-reliability assumptions plus the ad hoc test-construction assumption. The absence of invented entities is expected for an engineering paper, but the unstated hyperparameters and the biased test construction are the main burdens on the central claim.

free parameters (4)
  • Hard negative distance threshold α = not reported
    Equation (1) defines hard negatives by dist(f(x_i), f(x_j)) < α, but α is never specified. It controls the mining strategy and directly affects training.
  • Circle Loss margin m and scale γ = not reported
    Equation (2) introduces m and γ without values. These parameters control the decision boundary and hard-sample weighting.
  • Classification distance threshold = not reported
    Section 3.1 states a threshold is applied to decide positive vs negative pairs, but the threshold value is not reported, so the confusion matrix cannot be independently reproduced.
  • LoRA ranks r = 4, 8, 16
    Section 3.4 tests three ranks but does not report the selection criterion or the final rank used in the proposed model.
assumptions (4)
  • domain assumption Pretrained ViT/Unicom features trained on natural images transfer to neuron morphology patches.
    The entire fine-tuning approach assumes that ImageNet-scale features are a useful initialization for soma and fiber patterns. No domain-shift analysis is provided.
  • domain assumption The 273 manually annotated cross-modal neuron pairs are correct ground truth.
    Section 2.1 states pairs were manually annotated, but no inter-annotator agreement, alignment verification, or quality control is described.
  • ad hoc to paper The hard-negative test set selected by the model's own similarity scores yields an unbiased evaluation.
    Section 3.1 constructs 106 of 212 negatives as the highest-similarity non-matching pairs according to the model, making the test distribution a function of the model being evaluated. This is not standard independent evaluation.
  • domain assumption Two-photon and fMOST images of the same neuron depict the same soma and surrounding fibers at comparable scale.
    All positive pairs assume this correspondence. No explicit registration or scale normalization beyond manual annotation is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Few-Shot Metric Learning Method with Dual-Channel Attention for Cross-Modal Same-Neuron Identification." pith.science (2026). https://pith.science/paper/TM44R4CZ

@misc{pith2026250416520,
  author       = {Pith},
  title        = {Pith review of: A Few-Shot Metric Learning Method with Dual-Channel Attention for Cross-Modal Same-Neuron Identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TM44R4CZ}},
  note         = {Machine review of arXiv:2504.16520}
}
read the original abstract

In neuroscience research, achieving single-neuron matching across different imaging modalities is critical for understanding the relationship between neuronal structure and function. However, modality gaps and limited annotations present significant challenges. We propose a few-shot metric learning method with a dual-channel attention mechanism and a pretrained vision transformer to enable robust cross-modal neuron identification. The local and global channels extract soma morphology and fiber context, respectively, and a gating mechanism fuses their outputs. To enhance the model's fine-grained discrimination capability, we introduce a hard sample mining strategy based on the MultiSimilarityMiner algorithm, along with the Circle Loss function. Experiments on two-photon and fMOST datasets demonstrate superior Top-K accuracy and recall compared to existing methods. Ablation studies and t-SNE visualizations validate the effectiveness of each module. The method also achieves a favorable trade-off between accuracy and training efficiency under different fine-tuning strategies. These results suggest that the proposed approach offers a promising technical solution for accurate single-cell level matching and multimodal neuroimaging integration.

Figures

Figures reproduced from arXiv: 2504.16520 by the authors.

Figure 1
Figure 1. Network architecture and training strategy for metric learning. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Neuron retrieval results based on metric learning. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 4
Figure 4. Top-k accuracy comparison for different model variants. From the results, we observe that: The local-only model underperforms due to lack of global structural context; The global-only model lacks fine-grained detail needed for soma-level matching; Models using standard sampling or Triplet Loss fail to focus on hard-to-distinguish samples, leading [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Effects of Fine-Tuning Strategies on Neuron Similarity Accuracy. The results show that our proposed method consistently outperforms all baseline fine-tuning strategies across all Top-k evaluation metrics, demonstrating its superior capability in learning cross-modal fe…
Figure 6
Figure 6. Figure 6: Evaluation of Fine-Tuning Strategies: Similarity Heatmaps and Feature Distributions. From the visualization, it is evident that The original pretrained model fails to distinguish matched from unmatched pairs; Linear fine-tuning lacks sufficient adaptation due to minima…
Figure 7
Figure 7. Figure 7: Attention Visualization of the Dual-Channel Network (2) Visualization of Feature Space via Dimensionality Reduction To further assess the effectiveness of the proposed method in aligning cross-modal neuron features, we applied the t-distributed Stochastic Neighbor Embe…
Figure 9
Figure 9. Figure 9: Neuron similarity matrices and KDE curves using different methods. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 15 canonical work pages

  1. [4]

    M. Wang, K. Liu, J. Pan, J. Li, P. Sun, Y . Zhang, et al. Brain-wide projection reconstruction of single functionally defined neurons. Nature Communications, 2022, 13(1): 1531

    work page 2022

  2. [1]

    Introduction In neuroscience research, achieving cross-modal matching and integration at the single-neuron level is of critical importance for understanding the intrinsic relationship between neuronal function and structure. By integrating data at the single -cell level, functional and structural imaging can be combined to comprehensively characterize bot...

    work page 2024

  3. [2]

    Materials and methods 2.1 Dataset Construction for Neuron Similarity Learning To build a robust metric learning model for neuron similarity, we utilized manually annotated neuron pairs as the ground truth for both training and evaluation. A total of 273 cross-modal neuron pairs were collected and annotated, where each pair comprises corresponding neuron r...

    work page

  4. [3]

    Results 3.1 Evaluation of Method Effectiveness During the testing phase of the metric learning model, the evaluation samples include both matched neuron pairs—manually verified across two-photon and fMOST modalities—and unmatched pairs. As the number of negative ( unmatched) samples significantly exceeds that of positive samples, directly including all un...

    work page

  5. [5]

    X. Li, Y . Du, J.-F. Huang, W.-W. Li, W. Song, R.-N. Fan, et al. Link brain-wide projectome to neuronal dynamics in the mouse brain. Neuroscience Bulletin, 2024, 40(11): 1621-1634

    work page 2024

  6. [6]

    W. Zhou, S. Ke, W. Li, J. Yuan, X. Li, R. Jin, et al. Mapping the function of whole -brain projection at the single neuron level. Advanced Science, 2022, 9(33): e2202553

    work page 2022

  7. [7]

    W. Li, W. Chen, Z. Dai, X. Chai, S. An, Z. Guan, et al. Graph -based cell pattern recognition for merging the multi -modal optical microscopic image of neurons. Computer Methods and Programs in Biomedicine, 2024, 256: 108392

    work page 2024

  8. [8]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv, 2020, 2010.11929

    arXiv 2020

Show all 29 references
  1. [9]

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, et al. Swin transformer: Hierarchical vision transformer using shifted windows. in: IEEE/CVF International Conference on Computer Vision (ICCV). Montreal, Canada, 11-17 Oct. 2021, Proceedings of the IEEE, 2021: 9992-10002

    2021

  2. [10]

    Z. Liu, H. Hu, Y . Lin, Z. Yao, Z. Xie, Y . Wei, et al. Swin transformer v2: Scaling up capacity and resolution. in: IEEE/CVF Conference on Computer Vision Pattern Recognition (CVPR). Nashville, TN, USA, 11-17 Oct. 2021, Proceedings of the IEEE, 2021: 11999-12009

    2021

  3. [11]

    K. He, X. Chen, S. Xie, Y . Li, P. Doll'ar, R. B. Girshick. Masked autoencoders are scalable vision learners. in: IEEE/CVF Conference on Computer Vision Pattern Recognition (CVPR). Nashville, TN, USA, 19-25 Jun. 2021, Proceedings of the IEEE, 2021: 15979-15988

    2021

  4. [12]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, et al. Learning transferable visual models from natural language supervision. arXiv, 2021, 2103.00020

    2021 arXiv

  5. [13]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, et al. Segment anything. arXiv, 2023, 2304.02643

    2023 arXiv

  6. [14]

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, et al. Sam 2: Segment anything in images and videos. arXiv, 2024, 2408.00714

    2024 arXiv

  7. [15]

    Caron, H

    M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, et al. Emerging properties in self-supervised vision transformers. arXiv, 2021, 2104.14294

    2021 arXiv

  8. [16]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, et al. Dinov2: Learning robust visual features without supervision. arXiv, 2023, 2304.07193

    2023 arXiv

  9. [17]

    M. Ding, B. Xiao, N. C. F. Codella , P. Luo, J. Wang, L. Yuan. Davit: Dual attention vision transformers. in: European Conference on Computer Vision (ECCV). Tel -Aviv, 25-27 Oct. 2022, Springer, Cham, 2022

    2022

  10. [18]

    Carion, F

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, S. Zagoruyko. End -to-end object detection with transformers. arXiv, 2020, abs/2005.12872

    2020 arXiv

  11. [19]

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, J. Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv, 2020, abs/2010.04159

    2020 arXiv

  12. [20]

    Xiong, Z

    X. Xiong, Z. Wu, S. Tan, W. Li, F. Tang, Y . Chen, et al. Sam2-unet: Segment anything 2 makes strong encoder for natural and medical image segmentation. arXiv, 2024, abs/2408.08870

    2024

  13. [21]

    Z. Zhao, Y . Liu, H. Wu, Y . Li, S. Wang, L. Teng, et al. Clip in medical imaging: A comprehensive survey. arXiv, 2023, abs/2312.07353

    2023 arXiv

  14. [22]

    J. Zhu, Y . Qi, J. Wu. Medical sam 2: Segment medical images as video via segment anything model 2. arXiv, 2024, abs/2408.00874

    2024 arXiv

  15. [23]

    Ermolov, L

    A. Ermolov, L. Mirvakhabova, V . Khrulkov, N. Sebe, I. Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. arXiv, 2022, 2203.10833

    2022 arXiv

  16. [24]

    X. An, J. Deng, K. Yang, J. Li, Z. Feng, J. Guo, et al. Unicom: Universal and compact representation learning for image retrieval. arXiv, 2023, 2304.05884

    2023 arXiv

  17. [25]

    X. Wang, X. Han , W. Huang, D. Dong, M. R. Scott. Multi -similarity loss with general pair weighting for deep metric learning. in: IEEE/CVF Conference on Computer Vision Pattern Recognition (CVPR). Long Beach, CA, USA June 16, Proceedings of the IEEE, 2019: 5017-5025

    2019

  18. [26]

    Y . Sun, C. Cheng, Y . Zhang, C. Zhang, L. Zheng, Z. Wang, et al. Circle loss: A unified perspective of pair similarity optimization. in: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 14, Proceedings of the IEEE, 2020: 6397-6406

    2020

  19. [27]

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

    2021 arXiv

  20. [28]

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra. Grad -cam: Visual explanations from deep networks via gradient-based localization. in: IEEE International Conference on Computer Vision (ICCV). Venice, Italy, 22-29 Oct. 2017, Proceedings of the IEEE, 2017...

    2017

  21. [29]

    Van der Maaten, G

    L. Van der Maaten, G. Hinton. Visualizing data using t -sne. Journal of machine learning research, 2008, 9(11)

    2008

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.