pith. sign in

arxiv: 2504.00470 · v2 · pith:BIDSTQQXnew · submitted 2025-04-01 · 💻 cs.LG · cs.CV

Less is More: Efficient Black-box Attribution via Minimal Interpretable Subset Selection

Ruoyu Chen , Siyuan Liang , Jingzhi Li , Shiming Liu , Li Liu , Hua Zhang , Xiaochun Cao This is my paper

Pith reviewed 2026-05-22 21:26 UTC · model grok-4.3

classification 💻 cs.LG cs.CV
keywords black-box attributionsubmodular subset selectioninterpretable regionsmodel explanationbidirectional greedy searchfoundation modelsinsertion deletion evaluation
0
0 comments X

The pith

Reformulating black-box attribution as submodular subset selection identifies key input regions more faithfully using fewer samples.

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

The paper addresses the combinatorial explosion when attributing model decisions to discrete inputs such as image regions by recasting the task as an optimization problem that selects minimal important subsets. It introduces LiMA, which defines a submodular function to score how subsets influence predictions and applies a bidirectional greedy search to rank regions efficiently while locating both high- and low-impact areas. A sympathetic reader would care because the method promises explanations that remain accurate yet require less input data, backed by tests on eight foundation models that report gains on standard faithfulness measures and faster runtime than plain greedy search.

Core claim

LiMA reformulates the attribution of important regions as an optimization problem for submodular subset selection. First, a submodular function is designed to quantify subset importance and capture their impact on decision outcomes. Then, a bidirectional greedy search algorithm efficiently ranks input sub-regions by importance, identifying both the most and least important samples while ensuring an optimal attribution boundary that minimizes errors.

What carries the argument

The submodular function that quantifies subset importance together with the bidirectional greedy search algorithm for ranking and selecting minimal interpretable input regions.

If this is right

  • Provides faithful interpretations with fewer regions across eight foundation models.
  • Achieves an average 36.3 percent improvement in Insertion and 39.6 percent in Deletion metrics.
  • Runs 1.6 times faster than naive greedy search for attribution.
  • Yields 86.1 percent higher average highest confidence when explaining reasons for model prediction errors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The bidirectional search might be adapted to locate minimal subsets that preserve or remove specific model behaviors for targeted debugging.
  • If the submodular scoring generalizes, the same machinery could apply to other discrete inputs such as tokenized text sequences.
  • Identifying least-important regions could support data pruning experiments that test whether removing them leaves model accuracy intact.

Load-bearing premise

The submodular function designed to quantify subset importance accurately captures the true impact on decision outcomes without requiring post-hoc tuning or data-specific adjustments.

What would settle it

On a new set of foundation models or image datasets, the insertion and deletion faithfulness scores of LiMA would fail to exceed those of prior attribution methods while using fewer regions.

Figures

Figures reproduced from arXiv: 2504.00470 by Hua Zhang, Jingzhi Li, Li Liu, Ruoyu Chen, Shiming Liu, Siyuan Liang, Xiaochun Cao.

Figure 2
Figure 2. Figure 2: The framework of the proposed LIMA method. We begin by performing semantic sub-region division on the image, either using superpixel￾based methods or the Segment Anything algorithm. Next, we apply a bidirectional greedy search algorithm along with a designed submodular function to simultaneously identify the most and least important samples, ranking these sub-regions accordingly. Finally, based on sub-regi… view at source ↗
Figure 3
Figure 3. Figure 3: Statistics of strong interactive response times. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual explanations of the CLIP model using various attribution mechanisms, with our approach effectively reducing noise and eliminating [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Attribution visualizations of decision results for different multimodal foundation models on the ImageNet dataset. The first row shows the [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Attribution visualizations for audio classification on ImageBind [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of the method for discovering what causes foundation model prediction errors. The Insertion curve shows the correlation [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of the method for discovering what causes medical [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: An example of attribution results. A. Search region at the Insertion curve inflection point. B. Visualization using our strategy. C. Visualization using the baseline. 6.5.4 Ablation on Division Sub-region Number The sub-region division algorithm plays a key role in determining the quality of the search space elements. In addition to the choice of algorithm, the number of sub-regions, denoted as |V |, is a… view at source ↗
Figure 11
Figure 11. Figure 11: Additional interpretation visualization for different multimodal foundation models on the ImageNet dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional interpretation visualization for CLIP (ResNet-101) on the ImageNet dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Additional interpretation visualization for single-modal ResNet-101 on the ImageNet dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Additional interpretation visualization for single-modal vision transformer (Large) and swin-transformer (Large) on the ImageNet dataset. [PITH_FULL_IMAGE:figures/full_fig_p022_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Additional interpretation visualization for single-modal vision mamba (base) and mambavision (L2) on the ImageNet dataset. [PITH_FULL_IMAGE:figures/full_fig_p022_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Visualization of the method for discovering what causes model prediction errors on the CUB-200-2011 dataset. The first row shows the [PITH_FULL_IMAGE:figures/full_fig_p022_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Visualization of the method for discovering what causes model prediction errors on the CUB-200-2011 dataset. The first row shows the [PITH_FULL_IMAGE:figures/full_fig_p023_17.png] view at source ↗
read the original abstract

To develop a trustworthy AI system, which aim to identify the input regions that most influence the models decisions. The primary task of existing attribution methods lies in efficiently and accurately identifying the relationships among input-prediction interactions. Particularly when the input data is discrete, such as images, analyzing the relationship between inputs and outputs poses a significant challenge due to the combinatorial explosion. In this paper, we propose a novel and efficient black-box attribution mechanism, LiMA (Less input is More faithful for Attribution), which reformulates the attribution of important regions as an optimization problem for submodular subset selection. First, to accurately assess interactions, we design a submodular function that quantifies subset importance and effectively captures their impact on decision outcomes. Then, efficiently ranking input sub-regions by their importance for attribution, we improve optimization efficiency through a novel bidirectional greedy search algorithm. LiMA identifies both the most and least important samples while ensuring an optimal attribution boundary that minimizes errors. Extensive experiments on eight foundation models demonstrate that our method provides faithful interpretations with fewer regions and exhibits strong generalization, shows an average improvement of 36.3% in Insertion and 39.6% in Deletion. Our method also outperforms the naive greedy search in attribution efficiency, being 1.6 times faster. Furthermore, when explaining the reasons behind model prediction errors, the average highest confidence achieved by our method is, on average, 86.1% higher than that of state-of-the-art attribution algorithms. The code is available at https://github.com/RuoyuChen10/LIMA.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes LiMA, a black-box attribution method that reformulates identifying influential input regions as a submodular subset selection optimization problem. It introduces a custom submodular function to quantify subset importance for model decisions and a bidirectional greedy search algorithm to efficiently find both most- and least-important regions while defining an optimal attribution boundary. Experiments on eight foundation models report average gains of 36.3% on Insertion and 39.6% on Deletion metrics versus baselines, 1.6x speedup over naive greedy search, and 86.1% higher confidence on error explanations, with code released.

Significance. If the submodular function is verifiably monotone submodular and the empirical gains are robust, the approach could advance efficient, faithful black-box explanations for discrete inputs like images by using fewer regions and providing both positive and negative attributions. The code release and multi-model evaluation are positive factors supporting reproducibility and generalization claims.

major comments (3)
  1. [§3 (submodular function definition)] The design of the submodular function (abstract and §3) is asserted to quantify subset importance and capture decision impact, yet no formal proof of monotonicity or the diminishing-returns property is supplied, nor is there empirical verification on the model output surface. This is load-bearing because the bidirectional greedy algorithm's (1-1/e) approximation guarantee depends on it; without verification the reported metric gains and 'optimal attribution boundary' become purely empirical rather than theoretically supported.
  2. [§4 (experiments)] Experimental results (abstract and §4) report average improvements of 36.3% Insertion / 39.6% Deletion and 1.6x speedup without error bars, standard deviations, or statistical significance tests across the eight models. This weakens the strength of the generalization and efficiency claims.
  3. [§4 (experiments and algorithm)] No ablation is presented on the bidirectional search versus standard greedy or on how parameters of the submodular function were selected (abstract states 'design a submodular function' but provides no tuning details or sensitivity analysis). This leaves open whether the gains are driven by the specific search procedure or by implicit data-specific adjustments.
minor comments (2)
  1. [Abstract] Abstract contains grammatical issues ('which aim to identify' should be 'which aims to identify'; repeated 'on average' in the error-explanation sentence).
  2. [§3] Notation for the submodular function and the bidirectional search steps could be clarified with explicit pseudocode or a small worked example to improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We appreciate the referee's thorough review and valuable suggestions for improving the theoretical and empirical aspects of our work. Below, we provide point-by-point responses to the major comments.

read point-by-point responses

  1. Referee: [§3 (submodular function definition)] The design of the submodular function (abstract and §3) is asserted to quantify subset importance and capture decision impact, yet no formal proof of monotonicity or the diminishing-returns property is supplied, nor is there empirical verification on the model output surface. This is load-bearing because the bidirectional greedy algorithm's (1-1/e) approximation guarantee depends on it; without verification the reported metric gains and 'optimal attribution boundary' become purely empirical rather than theoretically supported.

    Authors: We recognize that the absence of a formal proof for the monotonicity and diminishing returns properties of our submodular function, as well as empirical verification, limits the theoretical support for the approximation guarantee. We will revise Section 3 to include a formal mathematical proof establishing these properties and add empirical analysis verifying the submodular behavior on the model outputs. This will strengthen the connection between the algorithm's guarantees and the reported performance improvements. revision: yes

  2. Referee: [§4 (experiments)] Experimental results (abstract and §4) report average improvements of 36.3% Insertion / 39.6% Deletion and 1.6x speedup without error bars, standard deviations, or statistical significance tests across the eight models. This weakens the strength of the generalization and efficiency claims.

    Authors: We agree that reporting without error bars or statistical tests reduces the robustness of our claims. In the revised version, we will update the experimental results in Section 4 to include error bars, standard deviations across the eight models, and statistical significance tests to validate the average improvements of 36.3% on Insertion and 39.6% on Deletion, as well as the speedup. revision: yes

  3. Referee: [§4 (experiments and algorithm)] No ablation is presented on the bidirectional search versus standard greedy or on how parameters of the submodular function were selected (abstract states 'design a submodular function' but provides no tuning details or sensitivity analysis). This leaves open whether the gains are driven by the specific search procedure or by implicit data-specific adjustments.

    Authors: We will incorporate ablations in the revised manuscript to compare the bidirectional greedy search against the standard greedy algorithm, quantifying the efficiency benefits. Additionally, we will provide details on the selection of parameters for the submodular function and include a sensitivity analysis to demonstrate that the performance gains are not due to data-specific tuning. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper reformulates black-box attribution as a submodular subset selection optimization problem and introduces a custom submodular function plus bidirectional greedy search as explicit algorithmic contributions. Performance claims (36.3% Insertion / 39.6% Deletion gains, 1.6x speed-up) are obtained by direct comparison against external baselines on eight foundation models rather than by algebraic reduction of fitted parameters or self-citations. No load-bearing step equates a derived quantity to its own inputs by construction, and the central optimization framework remains independent of the reported empirical outcomes.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the design of a submodular function that is assumed to quantify subset importance correctly and on the bidirectional greedy algorithm being able to locate an optimal attribution boundary. No explicit free parameters are named in the abstract, but the submodular scoring function itself functions as an implicit modeling choice whose form is not derived from first principles.

axioms (1)
  • domain assumption A submodular function can be defined that accurately quantifies the importance of input subsets for model decisions.
    Invoked when the paper states it designs a submodular function to assess interactions and capture impact on decision outcomes.

pith-pipeline@v0.9.0 · 5831 in / 1415 out tokens · 39048 ms · 2026-05-22T21:26:31.429562+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

97 extracted references · 97 canonical work pages · 3 internal anchors

  1. [1]

    A new metric based on association rules to assess feature- attribution explainability techniques for time series forecasting,

    A. Troncoso-Garc ´ıa, M. Mart´ınez-Ballesteros, F. Mart´ınez-´Alvarez, and A. Troncoso, “A new metric based on association rules to assess feature- attribution explainability techniques for time series forecasting,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1

    work page 2025

  2. [2]

    Towards human-centered ex- plainable ai: A survey of user studies for model explanations,

    Y . Rong, T. Leemann, T.-T. Nguyen, L. Fiedler, P. Qian, V . Unhelkar, T. Seidel, G. Kasneci, and E. Kasneci, “Towards human-centered ex- plainable ai: A survey of user studies for model explanations,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 46, no. 4, pp. 2104–2122, 2024. 1

    work page 2024

  3. [3]

    Explainable deep learning methods in medical image classification: A survey,

    C. Patr ´ıcio, J. C. Neves, and L. F. Teixeira, “Explainable deep learning methods in medical image classification: A survey,” ACM Computing Surveys, vol. 56, no. 4, pp. 1–41, 2023. 1

    work page 2023

  4. [4]

    Explainable Artificial Intelligence (XAI): Concepts, taxonomies, op- portunities and challenges toward responsible ai,

    A. B. Arrieta, N. D ´ıaz-Rodr´ıguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garc ´ıa, S. Gil-L ´opez, D. Molina, R. Benjamins et al. , “Explainable Artificial Intelligence (XAI): Concepts, taxonomies, op- portunities and challenges toward responsible ai,” Information fusion , vol. 58, pp. 82–115, 2020. 1 16

    work page 2020

  5. [5]

    Sim2Word: Explaining similarity with representative attribute words via counter- factual explanations,

    R. Chen, J. Li, H. Zhang, C. Sheng, L. Liu, and X. Cao, “Sim2Word: Explaining similarity with representative attribute words via counter- factual explanations,” ACM Transactions on Multimedia Computing, Communications and Applications, vol. 19, no. 6, pp. 1–22, 2023. 1

    work page 2023

  6. [6]

    Going beyond XAI: A systematic survey for explanation-guided learning,

    Y . Gao, S. Gu, J. Jiang, S. R. Hong, D. Yu, and L. Zhao, “Going beyond XAI: A systematic survey for explanation-guided learning,”ACM Computing Surveys, vol. 56, no. 7, pp. 1–39, 2024. 1

    work page 2024

  7. [7]

    End-to-end autonomous driving: Challenges and frontiers,

    L. Chen, P. Wu, K. Chitta, B. Jaeger, A. Geiger, and H. Li, “End-to-end autonomous driving: Challenges and frontiers,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 46, no. 12, pp. 10 164– 10 183, 2024. 1

    work page 2024

  8. [8]

    Less is more: Fewer interpretable region via submodular subset selection,

    R. Chen, H. Zhang, S. Liang, J. Li, and X. Cao, “Less is more: Fewer interpretable region via submodular subset selection,” in ICLR, 2024. 1, 2, 3, 4, 5, 9, 10, 11, 12, 14, 19, 22

    work page 2024

  9. [9]

    Making sense of dependence: Ef- ficient black-box explanations using dependence measure,

    P. Novello, T. Fel, and D. Vigouroux, “Making sense of dependence: Ef- ficient black-box explanations using dependence measure,” in NeurIPS, 2022, pp. 4344–4357. 1, 3, 9, 10, 11, 12, 13, 14, 19, 21, 22, 23

    work page 2022

  10. [10]

    Unifying fourteen post-hoc attribution methods with tay- lor interactions,

    H. Deng, N. Zou, M. Du, W. Chen, G. Feng, Z. Yang, Z. Li, and Q. Zhang, “Unifying fourteen post-hoc attribution methods with tay- lor interactions,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 7

    work page 2024

  11. [11]

    Interpreting object-level foundation models via visual precision search,

    R. Chen, S. Liang, J. Li, S. Liu, M. Li, Z. Huang, H. Zhang, and X. Cao, “Interpreting object-level foundation models via visual precision search,” arXiv preprint arXiv:2411.16198, 2024. 1

    work page arXiv 2024

  12. [12]

    Illuminating salient contributions in neuron activation with attribution equilibrium,

    W.-J. Nam and S.-W. Lee, “Illuminating salient contributions in neuron activation with attribution equilibrium,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 2, pp. 1120–1131, 2025. 1

    work page 2025

  13. [13]

    Local interpretations for explainable natural language processing: A survey,

    S. Luo, H. Ivison, S. C. Han, and J. Poon, “Local interpretations for explainable natural language processing: A survey,” ACM Computing Surveys, vol. 56, no. 9, pp. 1–36, 2024. 1

    work page 2024

  14. [14]

    A review and benchmark of feature importance methods for neural networks,

    H. Mandler and B. Weigand, “A review and benchmark of feature importance methods for neural networks,” ACM Computing Surveys , vol. 56, no. 12, pp. 1–30, 2024. 1

    work page 2024

  15. [15]

    Explainable ai (xai): Core ideas, techniques, and solutions,

    R. Dwivedi, D. Dave, H. Naik, S. Singhal, R. Omer, P. Patel, B. Qian, Z. Wen, T. Shah, G. Morgan et al. , “Explainable ai (xai): Core ideas, techniques, and solutions,” ACM Computing Surveys, vol. 55, no. 9, pp. 1–33, 2023. 1

    work page 2023

  16. [16]

    On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation,

    S. Bach, A. Binder, G. Montavon, F. Klauschen, K.-R. M ¨uller, and W. Samek, “On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation,” PloS One , vol. 10, no. 7, p. e0130140, 2015. 1, 3

    work page 2015

  17. [17]

    Axiomatic attribution for deep networks,

    M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic attribution for deep networks,” in ICML, 2017, pp. 3319–3328. 1, 3, 9, 10, 11, 12, 13, 18

    work page 2017

  18. [18]

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

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual explanations from deep networks via gradient-based localization,” International Journal of Computer Vision , vol. 128, no. 2, pp. 336–359, 2020. 1, 3, 11, 12

    work page 2020

  19. [19]

    iGOS++: integrated gradient optimized saliency by bilateral perturbations,

    S. Khorram, T. Lawson, and L. Fuxin, “iGOS++: integrated gradient optimized saliency by bilateral perturbations,” in CHIL, 2021, pp. 174–

    work page 2021

  20. [20]

    Gradient- based visual explanation for transformer-based clip,

    C. Zhao, K. Wang, X. Zeng, R. Zhao, and A. B. Chan, “Gradient- based visual explanation for transformer-based clip,” in ICML, 2024, pp. 61 072–61 091. 1, 3, 9, 10, 12, 13

    work page 2024

  21. [21]

    A unified approach to interpreting model predictions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” in NeurIPS, 2017, pp. 4765–4774. 1, 3, 10, 11, 12, 13, 14

    work page 2017

  22. [22]

    Explain any concept: Segment anything meets concept-based explanation,

    A. Sun, P. Ma, Y . Yuan, and S. Wang, “Explain any concept: Segment anything meets concept-based explanation,” in NeurIPS, 2023. 1, 3, 10, 11, 12, 13

    work page 2023

  23. [23]

    RISE: Randomized input sampling for explanation of black-box models,

    V . Petsiuk, A. Das, and K. Saenko, “RISE: Randomized input sampling for explanation of black-box models,” in BMVC, 2018, p. 151. 1, 3, 5, 9, 10, 11, 12, 13, 14, 19, 21

    work page 2018

  24. [24]

    Defining and extracting generalizable interaction primitives from dnns,

    L. Chen, S. Lou, B. Huang, and Q. Zhang, “Defining and extracting generalizable interaction primitives from dnns,” in ICLR, 2024. 1, 6, 7

    work page 2024

  25. [25]

    Towards the difficulty for a deep neural network to learn concepts of different complexities,

    D. Liu, H. Deng, X. Cheng, Q. Ren, K. Wang, and Q. Zhang, “Towards the difficulty for a deep neural network to learn concepts of different complexities,” in NeurIPS, 2023, pp. 41 283–41 304. 1

    work page 2023

  26. [26]

    Fujishige, Submodular functions and optimization

    S. Fujishige, Submodular functions and optimization. Elsevier, 2005. 1, 4, 6

    work page 2005

  27. [27]

    Learning transferable visual models from natural language supervision,

    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 ICML, 2021, pp. 8748–8763. 2, 8, 9, 10, 11, 12, 19

    work page 2021

  28. [28]

    ImageBind: One embedding space to bind them all,

    R. Girdhar, A. El-Nouby, Z. Liu, M. Singh, K. V . Alwala, A. Joulin, and I. Misra, “ImageBind: One embedding space to bind them all,” in CVPR, 2023, pp. 15 180–15 190. 2, 8, 9, 10, 12, 13, 19

    work page 2023

  29. [29]

    LanguageBind: Extending video-language pre- training to n-modality by language-based semantic alignment,

    B. Zhu, B. Lin, M. Ning, Y . Yan, J. Cui, W. HongFa, Y . Pang, W. Jiang, J. Zhang, Z. Li et al. , “LanguageBind: Extending video-language pre- training to n-modality by language-based semantic alignment,” in ICLR,

    work page

  30. [30]

    2, 8, 9, 10, 12, 13, 19

    work page

  31. [31]

    Quilt-1M: One million image- text pairs for histopathology,

    W. Ikezogwo, S. Seyfioglu, F. Ghezloo, D. Geva, F. Sheikh Mohammed, P. K. Anand, R. Krishna, and L. Shapiro, “Quilt-1M: One million image- text pairs for histopathology,” in NeurIPS, 2024, pp. 37 995–38 017. 2, 8, 13

    work page 2024

  32. [32]

    ImageNet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical image database,” inCVPR, 2009, pp. 248–255. 2, 8, 12

    work page 2009

  33. [33]

    VGGSound: A large- scale audio-visual dataset,

    H. Chen, W. Xie, A. Vedaldi, and A. Zisserman, “VGGSound: A large- scale audio-visual dataset,” in ICASSP, 2020, pp. 721–725. 2, 8, 12

    work page 2020

  34. [34]

    Caltech-UCSD Birds 200. Technical Report CNS-TR- 2010–001,

    P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Be-longie, and P. Perona, “Caltech-UCSD Birds 200. Technical Report CNS-TR- 2010–001,” Technical Report CNS-TR-2010–001, California Institute of Technology, 2010. 1, Tech. Rep., 2010. 2, 8, 12, 14

    work page 2010

  35. [35]

    Deep learning face attributes in the wild,

    Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes in the wild,” in ICCV, 2015, pp. 3730–3738. 2, 8

    work page 2015

  36. [36]

    VGGFace2: A dataset for recognising faces across pose and age,

    Q. Cao, L. Shen, W. Xie, O. M. Parkhi, and A. Zisserman, “VGGFace2: A dataset for recognising faces across pose and age,” in IEEE International Conference on Automatic Face and Gesture Recognition (FG), 2018, pp. 67–74. 2, 8

    work page 2018

  37. [37]

    Lung and colon cancer histopathological image dataset (LC25000),

    A. A. Borkowski, M. M. Bui, L. B. Thomas, C. P. Wilson, L. A. DeLand, and S. M. Mastorides, “Lung and colon cancer histopathological image dataset (LC25000),” arXiv preprint arXiv:1912.12142, 2019. 2, 8, 12

    work page arXiv 1912

  38. [38]

    Deep inside convolutional networks: visualising image classification models and saliency maps,

    K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep inside convolutional networks: visualising image classification models and saliency maps,” in ICLR (Workshop Poster), 2014. 3, 10, 11, 12

    work page 2014

  39. [39]

    Defining and extracting generalizable interaction primitives from dnns,

    L. Chen, S. Lou, B. Huang, and Q. Zhang, “Defining and extracting generalizable interaction primitives from dnns,” in ICLR, 2024. 3

    work page 2024

  40. [40]

    Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks,

    A. Chattopadhay, A. Sarkar, P. Howlader, and V . N. Balasubramanian, “Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks,” in WACV, 2018, pp. 839–847. 3, 14, 23

    work page 2018

  41. [41]

    Score-CAM: Score-weighted visual explanations for con- volutional neural networks,

    H. Wang, Z. Wang, M. Du, F. Yang, Z. Zhang, S. Ding, P. Mardziel, and X. Hu, “Score-CAM: Score-weighted visual explanations for con- volutional neural networks,” in CVPR Workshops, 2020, pp. 24–25. 3, 14

    work page 2020

  42. [42]

    ViT-CX: causal explanation of vision transformers,

    W. Xie, X.-H. Li, C. C. Cao, and N. L. Zhang, “ViT-CX: causal explanation of vision transformers,” in IJCAI, 2023, pp. 1569–1577. 3, 10, 12, 13

    work page 2023

  43. [43]

    A value for n-person games,

    L. S. Shapley, “A value for n-person games,” Annals of Mathematics Studies, vol. 28, pp. 307–318, 1953. 3

    work page 1953

  44. [44]

    Harsanyinet: Computing accurate shapley values in a single forward propagation,

    L. Chen, S. Lou, K. Zhang, J. Huang, and Q. Zhang, “Harsanyinet: Computing accurate shapley values in a single forward propagation,” in ICML, 2023, pp. 4804–4825. 3

    work page 2023

  45. [45]

    Problems with shapley-value-based explanations as feature importance measures,

    I. E. Kumar, S. Venkatasubramanian, C. Scheidegger, and S. Friedler, “Problems with shapley-value-based explanations as feature importance measures,” in ICML, 2020, pp. 5491–5500. 3, 7

    work page 2020

  46. [46]

    ”why should i trust you?

    M. T. Ribeiro, S. Singh, and C. Guestrin, “”why should i trust you?” explaining the predictions of any classifier,” inSIGKDD, 2016, pp. 1135–

    work page 2016

  47. [47]

    One explanation is not enough: structured attention graphs for image classification,

    V . Shitole, F. Li, M. Kahng, P. Tadepalli, and A. Fern, “One explanation is not enough: structured attention graphs for image classification,” in NeurIPS, 2021, pp. 11 352–11 363. 3

    work page 2021

  48. [48]

    Transformer interpretability beyond attention visualization,

    H. Chefer, S. Gur, and L. Wolf, “Transformer interpretability beyond attention visualization,” in CVPR, 2021, pp. 782–791. 3

    work page 2021

  49. [49]

    Vision transformers need registers,

    T. Darcet, M. Oquab, J. Mairal, and P. Bojanowski, “Vision transformers need registers,” in ICLR, 2024. 3

    work page 2024

  50. [50]

    Interpreting clip’s image representation via text-based decomposition,

    Y . Gandelsman, A. A. Efros, and J. Steinhardt, “Interpreting clip’s image representation via text-based decomposition,” in ICLR, 2024. 3

    work page 2024

  51. [51]

    Discover and cure: Concept-aware mitigation of spurious correlation,

    S. Wu, M. Yuksekgonul, L. Zhang, and J. Zou, “Discover and cure: Concept-aware mitigation of spurious correlation,” in ICML, 2023, pp. 37 765–37 786. 4

    work page 2023

  52. [52]

    Meaningfully debugging model mistakes using conceptual counterfactual explanations,

    A. Abid, M. Yuksekgonul, and J. Zou, “Meaningfully debugging model mistakes using conceptual counterfactual explanations,” in ICML, 2022, pp. 66–88. 4

    work page 2022

  53. [53]

    Over- looked factors in concept-based explanations: Dataset choice, concept learnability, and human capability,

    V . V . Ramaswamy, S. S. Kim, R. Fong, and O. Russakovsky, “Over- looked factors in concept-based explanations: Dataset choice, concept learnability, and human capability,” in CVPR, 2023, pp. 10 932–10 941. 4

    work page 2023

  54. [54]

    Talisman: targeted active learning for object detection with rare classes and slices using submodular mutual information,

    S. Kothawade, S. Ghosh, S. Shekhar, Y . Xiang, and R. Iyer, “Talisman: targeted active learning for object detection with rare classes and slices using submodular mutual information,” in ECCV, 2022, pp. 1–16. 4

    work page 2022

  55. [55]

    Ef- ficient modality selection in multimodal learning,

    Y . He, R. Cheng, G. Balasubramaniam, Y .-H. H. Tsai, and H. Zhao, “Ef- ficient modality selection in multimodal learning,” Journal of Machine Learning Research, vol. 25, no. 47, pp. 1–39, 2024. 4 17

    work page 2024

  56. [56]

    Marginal contribution feature importance-an axiomatic approach for explaining data,

    A. Catav, B. Fu, Y . Zoabi, A. L. W. Meilik, N. Shomron, J. Ernst, S. Sankararaman, and R. Gilad-Bachrach, “Marginal contribution feature importance-an axiomatic approach for explaining data,” in ICML, 2021, pp. 1324–1335. 4

    work page 2021

  57. [57]

    Stream- ing weak submodularity: Interpreting neural networks on the fly,

    E. Elenberg, A. G. Dimakis, M. Feldman, and A. Karbasi, “Stream- ing weak submodularity: Interpreting neural networks on the fly,” in NeurIPS, 2017, pp. 4044–4054. 4

    work page 2017

  58. [58]

    Learning to explain: An information-theoretic perspective on model interpretation,

    J. Chen, L. Song, M. Wainwright, and M. Jordan, “Learning to explain: An information-theoretic perspective on model interpretation,” in ICML, 2018, pp. 883–892. 4

    work page 2018

  59. [59]

    Scalable subset sampling with neural conditional poisson networks,

    A. Pervez, P. Lippe, and E. Gavves, “Scalable subset sampling with neural conditional poisson networks,” in ICLR, 2023, pp. 1–21. 4

    work page 2023

  60. [60]

    ”what data benefits my classifier?

    A. Chhabra, P. Li, P. Mohapatra, and H. Liu, “”what data benefits my classifier?” enhancing model performance and interpretability through influence-based data selection,” in ICLR, 2024. 4

    work page 2024

  61. [61]

    An analysis of approximations for maximizing submodular set functions—i,

    G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher, “An analysis of approximations for maximizing submodular set functions—i,” Mathe- matical programming, vol. 14, pp. 265–294, 1978. 4

    work page 1978

  62. [62]

    Lazier than lazy greedy,

    B. Mirzasoleiman, A. Badanidiyuru, A. Karbasi, J. V ondr ´ak, and A. Krause, “Lazier than lazy greedy,” in AAAI, 2015, pp. 1812–1818. 4, 6, 19

    work page 2015

  63. [63]

    Submodular batch selection for training deep neural networks,

    K. Joseph, K. Singh, and V . N. Balasubramanian, “Submodular batch selection for training deep neural networks,” in IJCAI, 2019, pp. 2677–

    work page 2019

  64. [64]

    Submodular functions, matroids, and certain polyhedra,

    J. Edmonds, “Submodular functions, matroids, and certain polyhedra,” Combinatorial Structures and Their Applications, pp. 69–87, 1970. 4

    work page 1970

  65. [65]

    SLIC superpixels compared to state-of-the-art superpixel methods,

    R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. S ¨usstrunk, “SLIC superpixels compared to state-of-the-art superpixel methods,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 34, no. 11, pp. 2274–2282, 2012. 5, 8

    work page 2012

  66. [66]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Dollar, and R. Girshick, “Segment anything,” in ICCV, 2023, pp. 4015–4026. 5, 8, 12, 23

    work page 2023

  67. [67]

    Unsupervised learning of visual representa- tions by solving jigsaw puzzles,

    M. Noroozi and P. Favaro, “Unsupervised learning of visual representa- tions by solving jigsaw puzzles,” in ECCV, 2016, pp. 69–84. 5

    work page 2016

  68. [68]

    Learn to threshold: Thresholdnet with confidence-guided manifold mixup for polyp segmentation,

    X. Guo, C. Yang, Y . Liu, and Y . Yuan, “Learn to threshold: Thresholdnet with confidence-guided manifold mixup for polyp segmentation,” IEEE Transactions on Medical Imaging , vol. 40, no. 4, pp. 1134–1146, 2020. 5

    work page 2020

  69. [69]

    Test-time prompt tuning for zero-shot generalization in vision-language models,

    M. Shu, W. Nie, D.-A. Huang, Z. Yu, T. Goldstein, A. Anandkumar, and C. Xiao, “Test-time prompt tuning for zero-shot generalization in vision-language models,” in NeurIPS, 2022, pp. 14 274–14 289. 5

    work page 2022

  70. [70]

    Handling open-set noise and novel target recognition in domain adaptive semantic segmentation,

    X. Guo, J. Liu, T. Liu, and Y . Yuan, “Handling open-set noise and novel target recognition in domain adaptive semantic segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 8, pp. 9846–9861, 2023. 5

    work page 2023

  71. [71]

    Cosface: Large margin cosine loss for deep face recognition,

    H. Wang, Y . Wang, Z. Zhou, X. Ji, D. Gong, J. Zhou, Z. Li, and W. Liu, “Cosface: Large margin cosine loss for deep face recognition,” in CVPR, 2018, pp. 5265–5274. 6

    work page 2018

  72. [72]

    Arcface: Additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,” in CVPR, 2019, pp. 4690–4699. 6, 8

    work page 2019

  73. [73]

    Submodularity in data subset selection and active learning,

    K. Wei, R. Iyer, and J. Bilmes, “Submodularity in data subset selection and active learning,” in ICML, 2015, pp. 1954–1963. 6

    work page 2015

  74. [74]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in CVPR, 2016, pp. 770–778. 8, 10, 11, 14, 19

    work page 2016

  75. [75]

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

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” in ICLR, 2021. 8, 11

    work page 2021

  76. [76]

    Swin transformer: Hierarchical vision transformer using shifted win- dows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted win- dows,” in ICCV, 2021, pp. 10 012–10 022. 8, 11, 19

    work page 2021

  77. [77]

    Vision mamba: Efficient visual representation learning with bidirectional state space model,

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: Efficient visual representation learning with bidirectional state space model,” in ICML, 2024. 8, 11, 19

    work page 2024

  78. [78]

    MambaVision: A hybrid mamba- transformer vision backbone,

    A. Hatamizadeh and J. Kautz, “MambaVision: A hybrid mamba- transformer vision backbone,” arXiv preprint arXiv:2407.08083 , 2024. 8, 11, 19

    work page arXiv 2024

  79. [79]

    MobileNetV2: Inverted residuals and linear bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “MobileNetV2: Inverted residuals and linear bottlenecks,” in CVPR, 2018, pp. 4510–4520. 8, 14

    work page 2018

  80. [80]

    EfficientNetV2: Smaller models and faster training,

    M. Tan and Q. Le, “EfficientNetV2: Smaller models and faster training,” in ICML, 2021, pp. 10 096–10 106. 8, 14

    work page 2021

Showing first 80 references.