Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that attribution for vision transformers is more faithful when an information bottleneck compresses several layers at once with one shared damping ratio, guaranteeing that discarded activations are unnecessary in every…

desk verdict A solid empirical multi-layer IBA extension whose advertised per-layer guarantee rests on an unproved and generally false monotonicity assumption. read the letter →

arxiv 2507.04388 v1 pith:6OHF7WEC submitted 2025-07-06 cs.CV

Jung-Ho Hong , Ho-Joong Kim , Kyu-Sung Jeon , Seong-Whan Lee This is my paper · ORCID
classification cs.CV
keywords featureattributioninformationbottleneckvisiontransformerinterpretabilitymutualvariationalboundinsertion/deletionfaithfulness
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 establish that attribution maps for vision transformers become more faithful when the information bottleneck is applied simultaneously to multiple layers rather than to one chosen layer, and that this can be done with a single shared per-token damping ratio. Existing single-layer information bottleneck attribution must be run once per layer, and no single layer dominates in explanation quality. CoIBA instead inserts bottlenecks into a range of layers and optimizes one universal damping ratio, so information needed at any targeted layer can compensate for over-compression in another. The central guarantee is that any activation discarded by the attribution is unnecessary in every targeted layer for the decision. If correct, this removes the heuristic search over layers while keeping computational cost comparable to a single information bottleneck run.

What carries the argument

The central object is the universal damping ratio λ = sigmoid(α) ∈ [0,1]^{P×1}, shared across all L bottleneck-inserted layers, with bottleneck variables Zl = λR′l + (1−λ)εl. Because λ is per-patch and channel-uniform, it measures token importance rather than per-neuron importance. The argument rides on the variational upper bound: assuming the mutual-information sequence I[Zl−1;Zl] is non-increasing under iterated noise injection, the inequality I[R1;Z1] ≥ (1/L) Σ I[Zl−1;Zl] bounds the average layer-wise compression by the first-layer compression, so the simplified objective max I[ZL;Y] − βI[R1;Z1] bounds every layer's compression with a single β. This upper bound is what converts the multi-layer objective into a tractable single-term objective and underpins the guarantee that discarded activations are unnecessary in every targeted layer.

What would settle it

Estimate the mutual information I[Zl−1;Zl] for each consecutive pair of CoIBA's bottleneck layers on a trained ViT-B/16, using the paper's own KL upper bound or a separate estimator, and check whether the sequence is non-increasing. A single violation, where some later pair exceeds an earlier pair, refutes the key inequality and the guarantee that discarded activations are unnecessary in every targeted layer.

Watch

Extended reading notes

Core claim

CoIBA claims that the relevant evidence for a vision transformer's decision is distributed across layers, and that a single-layer information bottleneck necessarily discards part of that evidence. By inserting bottleneck layers into a contiguous range of layers and sharing one universal damping ratio (one value per image patch, uniform across channels), the method estimates comprehensive relevant information; the shared ratio lets deeper layers' retained information compensate for over-compression in earlier layers. The objective replaces the sum of layer-wise compression terms with a single upper bound, leaving only one trade-off hyperparameter. With this bound, CoIBA guarantees that every activation it discards is unnecessary in every targeted layer, and the reported experiments show improved insertion/deletion, ROAD, and FunnyBirds scores over IBA and propagation-based baselines across ViT, DeiT, and Swin models.

Load-bearing premise

The guarantee rests on the unproven claim that injecting noise repeatedly makes the mutual information between consecutive bottleneck representations non-increasing, I[Zl;Zl+1] ≤ I[Zl−1;Zl]; if that monotonicity fails, the inequality does not bound the layer-wise compression terms and the simplified objective no longer delivers the per-layer guarantee.

Editorial extensions

If this is right

  • Attribution maps from CoIBA reflect evidence across a range of layers, so explanations no longer depend on a heuristic choice of a single bottleneck layer.
  • The variational upper bound removes the need to tune per-layer trade-off parameters; one universal damping ratio and one β suffice.
  • Because the compression term is bounded by first-layer information, the computation stays close to a single information bottleneck run despite compressing many layers.
  • If the guarantee holds, low-scored image regions can be treated as unnecessary in every targeted layer, strengthening the case for using the maps in failure diagnosis.
  • The reported insertion/deletion, ROAD, and FunnyBirds scores improve across ViT, DeiT, and Swin variants, including low-confidence (hard) samples.

Reading between the lines

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

  • The monotonicity assumption could be tested directly; if it fails only mildly, the method might still work in practice but the theoretical guarantee would need to be restated as approximate.
  • The universal per-token damping ratio effectively defines a token-level importance score, so λ itself could be repurposed as a coarse attribution map or used to select layers for deeper analysis.
  • The same multi-layer sharing scheme could apply to other transformer-based predictors, such as text or multimodal models, where evidence also accumulates across layers.
  • CoIBA's channel-uniform perturbation suggests a design principle for patch-based transformers: token-level masking can be more faithful than channel-level masking in information-bottleneck attribution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes CoIBA, a feature attribution method for vision transformers that inserts information bottlenecks into multiple layers and shares a single per-token damping ratio across them. The central claim is that minimizing a variational upper bound on the first-layer bottleneck mutual information compresses the information in every targeted layer, so that low-scored attributions are provably unnecessary for the decision. The experimental section reports substantial improvements over IBA and other ViT attribution baselines on insertion/deletion, ROAD, FunnyBirds, sensitivity-N, and sanity checks across ViT, DeiT, Swin, CLIP-pretrained, and self-supervised models.

Significance. If the theoretical claim were established, CoIBA would be a notable contribution: it promises a single-forward-pass attribution map that reflects evidence distributed across many layers while retaining the information-bottleneck-style guarantee of IBA. The empirical study is unusually broad, spanning multiple architectures, pretraining strategies, datasets, and evaluation protocols, and the authors provide open-source code. The method is likely practically useful as a heuristic. However, the load-bearing theoretical reduction from Eq. (4) to Eq. (8) is not proven, and the statement that CoIBA guarantees per-layer necessity is stronger than the derived inequality. The paper therefore needs major work before its central claim can be accepted.

major comments (3)
  1. [Sec. 3.4, Eq. (7)] The reduction from the layer-wise objective (4) to the simplified objective (8) rests on the asserted monotonicity I[Zl; Zl+1] ≤ I[Zl−1; Zl]. This is stated as following from 'iterated noise-injection' but no proof or empirical check is provided. The assertion is not true in general: with Zl = λR'_l + (1−λ)ε_l and R'_l a deterministic function of Zl−1, adding noise increases the conditional entropy H(Zl|Zl−1), but H(Zl) may also increase if intervening transformer blocks amplify the signal, so I[Zl−1; Zl] = H(Zl) − H(Zl|Zl−1) can rise. A simple scaling example (e.g., R'_2 = c Z1 with c ≫ 1 and small noise) yields I[Z1; Z2] > I[R1; Z1]. Since ViT residual and LayerNorm blocks are not information-contractive, nothing in the architecture enforces the needed ordering. Thus Eq. (7), and consequently the claimed guarantee, is not established. The supplementary derivation in App. A bounds only a single-layer KL term and does not address monotonicity across layers.
  2. [Sec. 3.4, Eq. (7) and Abstract] Even if Eq. (7) were true, it is an average bound: I[R1; Z1] ≥ (1/L) Σ_l I[Zl−1; Zl]. Minimizing β I[R1; Z1] bounds the sum of the layer-wise compression terms, not each term individually. The Abstract and Conclusion claim that CoIBA 'guarantees that the discarded activation is unnecessary in every targeted layer.' From the average bound one can only conclude that the average compression is controlled, not that every layer's discarded information is small. A per-layer upper bound of the form I[Zl−1; Zl] ≤ c I[R1; Z1] with a constant c is needed, and no such bound is derived. This gap is load-bearing because the paper's central theoretical selling point is per-layer relevance.
  3. [Sec. 4.1 and Supp. C.2, Tables A and B] The hyperparameters s (departure layer) and e (arrival layer) are selected on the basis of insertion/deletion scores on the same datasets used for final evaluation (Table B, Fig. 8), and the final reported numbers come from the best-performing settings. The trade-off β is also chosen using validation-set behavior (Supp. Table G). This creates a selection-on-evaluation-metric risk: part of the reported advantage may reflect tuning to the benchmark rather than a genuine property of the method. The authors should either use a separate validation split for hyperparameter selection, report results for a prespecified default, or present the sensitivity of the headline tables to the chosen configuration. They should also report error bars or multiple seeds, since all tables are point estimates.
minor comments (5)
  1. [Sec. 3.1, Eq. (2)] Use semicolon notation consistently in the mutual information terms: I[Rl; Zl] rather than I[Rl, Zl].
  2. [Sec. 3.2] The phrase 'demonstrate our insistence' should be 'demonstrate our claim' or 'demonstrate our motivation'.
  3. [Tables 1 and 2] The arrow notation in the table captions is easy to misread. Please state explicitly, e.g., 'higher is better for insertion and LeRF; lower is better for deletion and MoRF.'
  4. [Sec. 3.3] The term 'universal damping ratio' is used for a per-token vector shared across layers; 'shared' or 'common' would be less ambiguous than 'universal', which suggests a single scalar.
  5. [Page 1] There is a stray word 'CannonRacket' immediately after the Abstract; it appears to be an editing artifact and should be removed.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity in the CoIBA derivation; the reported gains are mildly inflated by hyperparameter selection on the evaluation metric, and Eq. (7) rests on an unproved monotonicity assumption (a correctness gap, not circularity).

  1. fitted input called prediction [Supplementary Material C.2 (Hyperparameters Selected in CoIBA), Tables A-B; main text Sec. 4.3]
    "We empirically select the hyperparameter, which broadly yields the best correctness performance. Tab. B illustrates the quantitative comparisons against various hyperparameter settings."

    The departure/arrival layer indices and trade-off parameter beta are chosen by comparing insertion/deletion 'correctness performance' on the same dataset (Tab. B), and the same metric is then reported in Table 1 as evidence that 'the attributions provided by CoIBA yield predominant correctness performance compared to baselines.' The reported comparison is therefore partly forced by the hyperparameter choice rather than being an independent prediction of faithfulness. The effect is mild because the core damping ratio lambda is learned from Eq. (8) via the information-bottleneck objective, not fitted to the insertion/deletion or ROAD metrics.

full rationale

The central derivation of CoIBA is not circular in the sense of reducing its outputs to its inputs. The shared damping ratio lambda is trained on the information-bottleneck objective (Eq. (8)), and the resulting attribution maps are evaluated post hoc with insertion/deletion, ROAD, FunnyBirds, and sanity checks; no attribution value is defined as the quantity being predicted. The paper does not rely on load-bearing self-citations to justify its method, and no ansatz is smuggled in via a citation. The one evaluation-adjacent concern is that the layer range and beta are selected by comparing the same insertion/deletion correctness scores that later appear as headline results (Supplementary Sec. C.2, Tab. B), which mildly favors CoIBA in Tables 1-2. This is a selection-bias issue rather than derivation circularity. The larger theoretical risk is not circularity but an unsupported claim: Eq. (7) requires the asserted monotonicity 'the iterated noise-injection procedure diminishes the mutual information among the internal and bottleneck representations, i.e., I[Zl;Zl+1] <= I[Zl-1;Zl]' (Sec. 3.4). No proof or empirical check is given, and transformer blocks are not information-contractive in general, so the reduction from Eq. (4) to Eq. (8) is not established. If Eq. (7) fails, minimizing beta*I[R1;Z1] bounds only an average of layer-wise compression terms and does not substantiate the abstract's 'guarantees ... in every targeted layer.' That is a correctness gap to be fixed, not a circularity.

Assumptions & free parameters 3 free parameters · 2 assumptions · 0 invented entities

The central claim rests on two hyperparameters and a domain assumption of monotonic mutual information decrease. The latter is the weakest link: without it, the variational upper bound is not established.

free parameters (3)
  • beta (trade-off parameter) = 1 for all models
    Selected based on insertion/deletion performance; see Sec. C.2 and Table B. Controls the balance between relevancy and compression in Eq. (8).
  • s (departure layer) = 4 for ViT-B/16, 8 for large models
    Empirically chosen to maximize faithfulness scores; see Table B.
  • e (arrival layer) = 12 for ViT-B/16, 24 for large models
    Empirically chosen to maximize faithfulness scores; see Table B.
assumptions (2)
  • ad hoc to paper I[Zl; Zl+1] <= I[Zl-1; Zl] for all targeted layers (monotonicity under noise injection)
    Stated in Sec. 3.4 as the basis for Eq. (7) but not proven; it is load-bearing for the variational upper bound.
  • domain assumption The prior P(Zl) can be replaced by Gaussian Q(Zl) = N(mu_Rl, sigma^2_Rl) such that the KL divergence overestimates the mutual information
    Standard variational approximation in deep VIB and IBA; used in Eq. (6).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers." pith.science (2026). https://pith.science/paper/6OHF7WEC

@misc{pith2026250704388,
  author       = {Pith},
  title        = {Pith review of: Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6OHF7WEC}},
  note         = {Machine review of arXiv:2507.04388}
}
read the original abstract

The feature attribution method reveals the contribution of input variables to the decision-making process to provide an attribution map for explanation. Existing methods grounded on the information bottleneck principle compute information in a specific layer to obtain attributions, compressing the features by injecting noise via a parametric damping ratio. However, the attribution obtained in a specific layer neglects evidence of the decision-making process distributed across layers. In this paper, we introduce a comprehensive information bottleneck (CoIBA), which discovers the relevant information in each targeted layer to explain the decision-making process. Our core idea is applying information bottleneck in multiple targeted layers to estimate the comprehensive information by sharing a parametric damping ratio across the layers. Leveraging this shared ratio complements the over-compressed information to discover the omitted clues of the decision by sharing the relevant information across the targeted layers. We suggest the variational approach to fairly reflect the relevant information of each layer by upper bounding layer-wise information. Therefore, CoIBA guarantees that the discarded activation is unnecessary in every targeted layer to make a decision. The extensive experimental results demonstrate the enhancement in faithfulness of the feature attributions provided by CoIBA.

Figures

Figures reproduced from arXiv: 2507.04388 by the authors.

Figure 1
Figure 1. The quantitative and qualitative comparisons of at￾tribution maps generated from different layers. We illustrate the visualized attribution maps from the first to fourth columns in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Quantitative and qualitative results of IBA*. In the first row, the horizontal line indicates the performance of layer￾specific relevant information, i.e., IBA. We report the difference in insertion and deletion scores (∆InsDel). The higher score is better for this metric. The visualized attribution maps are illustrated in the second row obtained from VIT-B-16/224 [14]. the overall procedure of CoIBA. Finally, we in… view at source ↗
Figure 4
Figure 4. The quantitative comparisons of explanation meth￾ods in high-confident (easy) and low-confident (hard) input samples. The confidence score is the softmax probability of the target class predicted by the model. We compare the information￾bottleneck-based methods including IBA [32], IBA∗ , and CoIBA (marked with a circle), and the propagation-based explanation baselines including Beyond [11], IIA [6], and Chefer-LRP [… view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: An overview of CoIBA. CoIBA restricts the information flow using a universal damping ratio while upper-bounding the infor￾mation of each layer. After that, CoIBA aggregates the relevant information from each layer to produce an attribution map. tween bottleneck represe…
Figure 6
Figure 6. Figure 6: Quantitative comprehensive assessment on FunnyBirds experiment: The abbreviations Com. Cor. and Con. indicate Completeness, correctness, and contrastivity. The center score indicates the comprehensive result of enumerated aspects. Acc. and B.I. indicate accuracy and ba…
Figure 7
Figure 7. Figure 7: Quantitative difficulty-aware correctness assessment on insertion/deletion and ROAD. We measure differences in inser￾tion/deletion (∆InsDel↑) and ROAD (∆ROAD↑) scores. We fill the regions including low-confident samples and high-confident samples with red and blue, res…
Figure 8
Figure 8. Figure 8: Comparisons of different departure s and arrival e layer settings. We plot the hyper-parameters selected for CoIBA as a solid line. We compare the quantitative results of the discrep￾ancy between AUC scores of insertion/deletion (∆InsDel) with different intervals of co…
Figure 9
Figure 9. Figure 9: Ablation study on layers to analyze universal damp￾ing ratio [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 34 canonical work pages

  1. [1]

    Quantifying attention flow in transformers

    Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , pages 4190–4197, 2020. 2

    work page 2020

  2. [2]

    Sanity checks for saliency maps

    Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Good- fellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. Advances in Neural Information Processing Systems, 31, 2018. 5

    work page 2018

  3. [3]

    Deep variational information bottleneck

    Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. In Inter- national Conference on Learning Representations, 2017. 2, 3

    work page 2017

  4. [4]

    Towards better understanding of gradient-based attri- bution methods for deep neural networks

    Marco Ancona, Enea Ceolini, Cengiz ¨Oztireli, and Markus Gross. Towards better understanding of gradient-based attri- bution methods for deep neural networks. In International Conference on Learning Representations, 2018. 3

    work page 2018

  5. [5]

    BEit: BERT pre-training of image transformers

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations, 2022. 3

    work page 2022

  6. [6]

    Visual explanations via iterated integrated attributions

    Oren Barkan, Yuval Asher, Amit Eshel, Noam Koenigstein, et al. Visual explanations via iterated integrated attributions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2073–2084, 2023. 1, 2, 4, 6, 7

    work page 2023

  7. [7]

    Layer-wise relevance propagation for neural networks with local renormalization layers

    Alexander Binder, Gr ´egoire Montavon, Sebastian La- puschkin, Klaus-Robert M ¨uller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In Artificial Neural Networks and Machine Learning–ICANN 2016: 25th International Conference on Artificial Neural Networks, Barcelona, Spain, September 6-9, 2016, ...

    work page 2016

  8. [8]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650–9660, 2021. 1, 3

    work page 2021

Show all 42 references
  1. [9]

    Generic attention- model explainability for interpreting bi-modal and encoder- decoder transformers

    Hila Chefer, Shir Gur, and Lior Wolf. Generic attention- model explainability for interpreting bi-modal and encoder- decoder transformers. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 397–406,

  2. [10]

    Transformer inter- pretability beyond attention visualization

    Hila Chefer, Shir Gur, and Lior Wolf. Transformer inter- pretability beyond attention visualization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 782–791, 2021. 1, 2, 4, 6, 7

    2021

  3. [11]

    Beyond intuition: Rethinking token attributions in- side transformers

    Jiamin Chen, Xuhong Li, Lei Yu, Dejing Dou, and Haoyi Xiong. Beyond intuition: Rethinking token attributions in- side transformers. Transactions on Machine Learning Re- search, 2022. 1, 2, 4, 6, 7

    2022

  4. [12]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009. 6

    2009

  5. [13]

    Towards a rigorous science of interpretable machine learning

    Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 1

    2017 arXiv

  6. [14]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on...

    2020

  7. [15]

    Eva: Exploring the limits of masked visual representa- tion learning at scale

    Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representa- tion learning at scale. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition,...

    2023

  8. [16]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000– 16009, 2022. 1, 3

    2022

  9. [17]

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

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kada- vath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robust- ness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF Internation...

    2021

  10. [18]

    Natural adversarial examples

    Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Stein- hardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 15262–15271, 2021. 6

    2021

  11. [19]

    Fun- nybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods

    Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Fun- nybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3981– 3991, 2023. 2, 6

    2023

  12. [20]

    Optimizing rele- vance maps of vision transformers improves robustness

    Idan Schwartz Hila Chefer and Lior Wolf. Optimizing rele- vance maps of vision transformers improves robustness. In Advances in Neural Information Processing Systems , 2022. 4, 5

    2022

  13. [21]

    A benchmark for interpretability methods in deep neural networks

    Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. Advances in Neural Information Process- ing Systems, 32, 2019. 7

    2019

  14. [22]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

    arXiv

  15. [23]

    Auto-encoding varia- tional bayes

    Diederik P Kingma and Max Welling. Auto-encoding varia- tional bayes. arXiv preprint arXiv:1312.6114, 2013. 3

    2013 arXiv

  16. [24]

    Similarity of neural network represen- tations revisited

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network represen- tations revisited. In International Conference on Machine Learning, pages 3519–3529. PMLR, 2019. 8

    2019

  17. [25]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012–10022, 2021. 6

    2021

  18. [26]

    Swin transformer v2: Scaling up capacity and resolution

    Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12009–12019, 2022. 6

    2022

  19. [27]

    Safe and interpretable machine learning: a methodological review

    Clemens Otte. Safe and interpretable machine learning: a methodological review. Computational Intelligence in Intel- ligent Data Analysis, pages 111–122, 2013. 1

    2013

  20. [28]

    Rise: Random- ized input sampling for explanation of black-box models

    Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Random- ized input sampling for explanation of black-box models. British Machine Vision Conference, 2018. 2, 4, 6

    2018

  21. [29]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In International Conference on Machine Learning...

    2021

  22. [30]

    A consistent and efficient eval- uation strategy for attribution methods

    Yao Rong, Tobias Leemann, Vadim Borisov, Gjergji Kas- neci, and Enkelejda Kasneci. A consistent and efficient eval- uation strategy for attribution methods. InInternational Con- ference on Machine Learning, pages 18770–18795. PMLR,

  23. [31]

    Towards explain- able artificial intelligence

    Wojciech Samek and Klaus-Robert M¨uller. Towards explain- able artificial intelligence. Explainable AI: interpreting, ex- plaining and visualizing deep learning , pages 5–22, 2019. 1

    2019

  24. [32]

    Restricting the flow: Information bottlenecks for attri- bution

    Karl Schulz, Leon Sixt, Federico Tombari, and Tim Land- graf. Restricting the flow: Information bottlenecks for attri- bution. In International Conference on Learning Represen- tations, 2020. 1, 2, 3, 4, 6, 7

    2020

  25. [33]

    Grad-cam: Visual explanations from deep networks via gradient-based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. InProceedings of the IEEE/CVF International Conference on Computer Vision , pages 618–...

    2017

  26. [34]

    How to train your vit? data, augmentation, and regularization in vision transformers

    Andreas Peter Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. Transactions on Machine Learning Re- search, 2024. 6

    2024

  27. [35]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International Conference on Machine Learning, pages 3319–3328. PMLR, 2017. 1

    2017

  28. [36]

    The information bottleneck method

    Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 2

    2000 arXiv

  29. [37]

    Training data-efficient image transformers & distillation through at- tention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through at- tention. In International Conference on Machine Learning , pages 10347–10357. PMLR, 2021. 4, 6

    2021

  30. [38]

    Deit iii: Revenge of the vit

    Hugo Touvron, Matthieu Cord, and Herv ´e J ´egou. Deit iii: Revenge of the vit. In Proceedings of the European Con- ference on Computer Vision, pages 516–533. Springer, 2022. 6

    2022

  31. [39]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Process- ing, 13(4):600–612, 2004. 2

    2004

  32. [40]

    Vit-cx: causal explanation of vision transformers

    Weiyan Xie, Xiao-Hui Li, Caleb Chen Cao, and Nevin L Zhang. Vit-cx: causal explanation of vision transformers. In Proceedings of the Thirty-Second International Joint Con- ference on Artificial Intelligence, pages 1569–1577, 2023. 1, 2, 6, 7

    2023

  33. [41]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975–11986, 2023. 1

    2023

  34. [42]

    Fine-grained neural net- work explanation by identifying input features with predic- tive information

    Yang Zhang, Ashkan Khakzar, Yawei Li, Azade Farshad, Seong Tae Kim, and Nassir Navab. Fine-grained neural net- work explanation by identifying input features with predic- tive information. Advances in Neural Information Process- ing Systems, 34:20040–20051, 2021. 2, 3, 5 Compr...

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.