Pith. sign in

REVIEW 3 major objections 5 minor 68 references

On the Performance of Concept Probing: The Influence of the Data (Extended Version)

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 ↗

Pith's one-line read Concept probes reach 97% of peak accuracy with as few as 250 labeled samples, and can reuse the original model's training data without losing accuracy.

desk verdict A genuinely useful empirical study of what data concept probing needs, with headline numbers that mostly hold up but are narrower than they look. read the letter →

arxiv 2507.18550 v1 pith:2624OZYT submitted 2025-07-24 cs.AI cs.CVcs.LGcs.NE

Manuel de Sousa Ribeiro , Afonso Leote , João Leite This is my paper
classification cs.AIcs.CVcs.LGcs.NE
keywords conceptprobingexplainableAIneuralnetworkinterpretabilityprobetrainingdatareuselabelnoiseGTSRBCUB
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 asks how the data used to train concept probes, classifiers that read a neural network's internal activations and predict human-defined concepts, shapes their accuracy. Probing eight image classifiers across six datasets with five probe architectures, it finds that concepts relevant to the model's task can be probed accurately with very little data: on average the probe reaches 97.3% of its maximum measured accuracy with only 250 training samples, roughly 0.8% of the original model's training data. It also reports that reusing the original model's training data leaves probe accuracy essentially unchanged, and that probes tolerate moderate label noise, with a 20% noise rate costing about 9.3% relative accuracy. The paper presents these results as evidence against the folk assumptions that probing requires large, fresh, high-quality datasets, a matter that matters because labeled concept data is usually the dominant cost of applying concept probing.

What carries the argument

The central object is the probe $g$, a classifier trained on activations $f_u(x)$ produced by a layer or unit set $u$ of an original model $f$ to predict a concept value $c$; the paper considers five probe architectures: linear Logistic and Ridge classifiers, LightGBM decision trees, a feedforward neural network, and a mapping-network probe that selects units from multiple layers via an input-reduce procedure. The experiments vary four dimensions of the probe dataset $D_g$, namely its size, the scale of the probed model, the fraction of samples reused from the original model's training set, and the amount of label noise, while separating concepts into relevant and non-relevant relative to the model's task because the authors hypothesize that task-relevant concepts are easier to probe. The paper also contributes new concept labels for GTSRB and revised attribute labels for CUB, which serve as test beds for the data-quality experiments.

What would settle it

Train the same five probes on randomly selected layers, including early convolution layers, and on additional datasets, then measure accuracy at 250 training samples and at 100% data reuse; if probe accuracy for relevant concepts falls well below 97.3% of its peak, or if full reuse produces a statistically significant accuracy drop, the paper's central claims would be contradicted.

Watch

Extended reading notes

Core claim

The central claim is that concept probing is data-cheap, scale-friendly, and reuse-tolerant for image classifiers. For task-relevant concepts, probe accuracy climbs quickly and stabilizes around 200 training samples, while non-relevant concepts need more data and perform substantially worse. Scaling the original model up within the tested ranges slightly improves probe accuracy, by an average of 1.8%, whereas scaling it down hurts accuracy by an average of 3.9%, suggesting that larger representations remain probeable. Probe accuracy stays essentially flat when the probe is trained entirely on the original model's own training data, with a best-fit slope of $-4.1\times10^{-3}$ that is not statistically significant. Label noise up to roughly 20% causes a 9.3% relative accuracy drop, and naturally noisy CUB attribute labels, where 18.2% of labels are incorrect and the noise is concentrated on false positives, cost 13.1% relative accuracy, implying that data quality still matters even though probes are moderately robust.

Load-bearing premise

The broad claims that larger models are easy to probe and that reusing the original training data is harmless rest on experiments with only two datasets, GTSRB and XTRAINS, and on layers chosen by inspection because earlier layers seemed not to capture the probed concepts; if those settings are atypical, the generality of those claims fails.

Editorial extensions

If this is right

  • Practitioners can build reliable probes for task-relevant concepts with a few hundred balanced labeled examples rather than large dedicated datasets.
  • Existing training data of the probed model can be repurposed for probing, easing application in data-scarce domains.
  • Probing remains feasible, and may even benefit slightly, when the original model is scaled up, so growing model size alone is not a barrier.
  • Probe results should be reported together with data-quality checks, because label noise above roughly 30% degrades probes sharply and realistic non-random noise is costlier than random noise.
  • The released GTSRB concept labels and revised CUB attribute labels give other researchers standard data for further probing experiments.

Reading between the lines

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

  • A testable extension left implicit by the paper is to repeat the data-size and reuse curves on transformer-based language models, since the current experiments cover only image classifiers.
  • Because the scaled-model and reuse experiments run only on GTSRB and XTRAINS, the strongest reading is that these effects hold for visual classifiers trained from scratch on moderately sized datasets; extrapolation to very large pretrained models is an inference rather than a result of this paper.
  • The paper's relevance-based distinction suggests a practical heuristic, that practitioners should measure how separable and task-relevant a concept is before deciding how much probe data to collect, although this heuristic is not itself tested here.
  • The finding that real CUB label noise is mostly false positives implies that simple noise-rate estimates can understate the true cost of noisy concept labels, so auditing which label class carries the noise would sharpen quality control.
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 investigates how characteristics of the data used to train concept probes affect probe performance in image classification. Four dimensions are studied: training-set size, original-model size, reuse of the original model's training data, and label noise. The authors use five probe architectures (logistic regression, ridge regression, LightGBM, a small neural network, and a mapping-network probe) on eight models spanning six datasets. The central empirical claims are that task-relevant concepts can be probed with few training samples (97.3% of the maximum measured accuracy with 250 samples), that larger original models do not hurt probe accuracy, that reusing the original model's training data does not degrade probe accuracy, and that probes tolerate moderate label noise. The paper also contributes relabeled CUB attributes and a GTSRB concept ontology with labels.

Significance. If the central claims hold, the paper gives practitioners concrete guidance: hundreds of labeled examples may suffice for probing relevant concepts, and existing training data can be repurposed without obvious harm. This addresses a real gap, since most concept-probing research has focused on probe architecture or interpretation rather than on data requirements. The paper's strengths include the breadth of probe architectures compared, the balanced repeated-trial protocol, the careful manual relabeling of CUB attributes in Appendix A, and the explicit ontology-based labeling procedure for GTSRB in Appendix B. The main weakness is that the strongest headline claims are built on a restricted set of hand-picked layers and, for the model-size and data-reuse conclusions, on only two datasets; these restrictions are disclosed but their effect on the generality of the claims is not quantified.

major comments (3)
  1. [Section 3.2, Probed Activations] The layer-selection procedure removes early layers by inspection: for ResNet models only the first and last convolution blocks of the fourth residual block are probed, and for ViT only the last transformer block, because the authors state that 'the activations resulting from the previous blocks seemed to not yet capture the probed concepts.' The headline result in Section 4 that probe accuracy reaches 97.3% of its maximum with 250 training samples is therefore an average over a set of late, hand-picked layers, and the excluded early layers never enter the average. A practitioner who must search over layers without knowing which one captures a concept faces a different data requirement. The disclosure of the selection is transparent, but its effect on the headline number is not quantified; please report per-layer results or explicitly qualify the claim as applying to late, concept-bearing layers.
  2. [Section 4, Effect of Train Dataset Size] The method behind the '97.3% of its maximum measured value with as few as 250 training samples' statistic is not defined. It is unclear whether the percentage is averaged over relevant concepts only, over all five probe architectures, and over all probed layers, and how 'maximum measured value' is computed from the curves in Figure 2. The figure shows only averaged curves over five repetitions with no error bars or confidence intervals. Since this number is the paper's headline claim and the basis for the practical recommendation, the exact aggregation protocol and a measure of variability must be provided so the claim can be independently verified.
  3. [Sections 5-6, Figures 3-4] The conclusions that larger original models are not harder to probe and that reusing the original model's training data does not affect probe accuracy are based on only the GTSRB and XTRAINS settings, one of which is synthetic. The paper acknowledges this limitation in Section 5, but Section 9 states these findings without qualification, for example 'repurposing the data involved in the training of the probed model to train probing models does not degrade their resulting test accuracy.' In addition, Figure 3 reports a 1.8% improvement and a 3.9% degradation without error bars or statistical tests, so the size-scaling effects could be within repetition noise. Please either restrict the conclusions to the two tested settings or provide additional evidence, and add variance reporting to Figure 3.
minor comments (5)
  1. [Section 4, parenthetical claim] The parenthetical 'about 0.8% of the original model's training data' is not a meaningful average across the six datasets, whose training-set sizes differ by orders of magnitude; please specify the dataset for which this holds or remove the parenthetical.
  2. [Figures 2-5] Since every result is averaged over five repetitions, the lack of error bars or per-repetition points makes it difficult to assess the stability of the reported differences; consider adding confidence intervals to all figures.
  3. [Section 3.1] The hyperparameter search is described only for the Ridge probe; please state explicitly how hyperparameters for the other probe architectures were chosen, even if they were fixed by default values.
  4. [Section 3.2] There is a typo in the XTRAINS description: 'PassangerTrains' should be 'PassengerTrains.'
  5. [Figure captions] The legends labeled 'Relevant Non-Relevant' in Figures 2-5 are not explained in the captions; please define 'relevant' and 'non-relevant' or point readers to Section 3.2.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all headline claims are empirical measurements of probe accuracy on held-out test data, not derivations from the inputs that define them.

full rationale

The paper's central claims concern measured probe accuracy as a function of training-data size (Section 4), original-model scale (Section 5), data reuse (Section 6), and label noise (Section 7). Each of these is an empirical curve computed on balanced held-out test sets D'_g, so the reported numbers are not derived from the probe's own training data or from the definitions of the concepts. The 97.3%-at-250-samples figure is a descriptive summary of measured accuracies, not a quantity forced by construction. The relevant/non-relevant distinction is defined via an ontology and a formal relevancy notion from the authors' prior work [13], but it is not defined in terms of probe accuracy; comparing measured accuracies across relevance classes is therefore not tautological. The paper does rely on author-generated resources (XTRAINS, the GTSRB ontology and labels, the CUB relabeling), and it cites the authors' own work for the relevancy definition and the dataset. However, these are experimental inputs, not load-bearing proof of the empirical conclusions: the probe accuracies are measured independently against fresh test examples. The layer-selection heuristic is also disclosed ('the activations resulting from the previous blocks seemed to not yet capture the probed concepts', Section 3.2), and it is a scope limitation about external validity rather than a circular reduction, since no layer choice is defined in terms of the probe accuracies it produces. The data-reuse analysis fits a line and reports a t-test; this is a descriptive regression, not a fitted parameter relabeled as a prediction. No equation in the paper reduces to its own input, and no uniqueness theorem or self-citation chain forces the stated conclusions. Hence the paper is empirically self-contained with respect to its claims, and no circular step is present.

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

No numerical parameters are fitted in this study; the paper contributes measurements. The claims rest on domain assumptions about what makes a concept 'relevant', on the correctness of the authors' manual relabeling of CUB attributes, and on the validity of balanced test accuracy as a probe-performance measure.

assumptions (4)
  • domain assumption A concept relevant to a model's task is likely reflected in its internal representations.
    Section 4 frames the data-size experiment around this expectation: 'if the concept C being probed is indeed relevant for the task... then its internal representations have likely distilled some information that is related to C.'
  • domain assumption Relevance can be determined from Description Logic ontologies built by the authors for GTSRB and XTRAINS, and by author judgment for ImageNet, DeepFashion, Places365, and CUB.
    Sections 3.2 and 6 classify concepts as relevant or non-relevant; for non-ontology datasets the paper states 'we believe the probed concepts are relevant' without a formal procedure.
  • domain assumption The manually relabeled CUB attribute labels are correct ground truth for measuring natural label noise.
    Appendix A describes the relabeling procedure and treats the new labels as authoritative; the 18.2% natural-noise figure and the corresponding 13.1% performance drop depend on this premise.
  • domain assumption Balanced accuracy on a held-out probe test set measures how much information representations carry about a concept.
    Section 2 states that accuracy of g on a balanced test set is generally considered, with higher accuracy indicating more information; this is the standard concept-probing assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Performance of Concept Probing: The Influence of the Data (Extended Version)." pith.science (2026). https://pith.science/paper/2624OZYT

@misc{pith2026250718550,
  author       = {Pith},
  title        = {Pith review of: On the Performance of Concept Probing: The Influence of the Data (Extended Version)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2624OZYT}},
  note         = {Machine review of arXiv:2507.18550}
}
read the original abstract

Concept probing has recently garnered increasing interest as a way to help interpret artificial neural networks, dealing both with their typically large size and their subsymbolic nature, which ultimately renders them unfeasible for direct human interpretation. Concept probing works by training additional classifiers to map the internal representations of a model into human-defined concepts of interest, thus allowing humans to peek inside artificial neural networks. Research on concept probing has mainly focused on the model being probed or the probing model itself, paying limited attention to the data required to train such probing models. In this paper, we address this gap. Focusing on concept probing in the context of image classification tasks, we investigate the effect of the data used to train probing models on their performance. We also make available concept labels for two widely used datasets.

Figures

Figures reproduced from arXiv: 2507.18550 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Probe performance by amount of training data. residual layers, we consider the activations of every fourth residual layer. For the ResNet50 (fP laces, fCUB) and ResNet101 (fF ashion) models, which are composed of 4 residual blocks, we probe the acti￾vations produced by the first and last convolution blocks of the fourth residual block, as the activations resulting from the previous blocks seemed to not yet capture t… view at source ↗
Figure 3
Figure 3. Probe performance by scaling of the original model’s size. results. Furthermore, in the case of MapNN, a t-test shows that its ac￾curacy is statistically significantly higher than the remaining probes when probing from fImageNet, fF ashion, fCUB, fGT SRB, fA, fB, and fC , with the maximum amount of available data. Additionally, Ridge seems to be the probe that is benefiting less from additional data, showing, on ave… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Probe performance by percentage of data reuse from the original model’s training data. If reusing the input samples x (i) from the original model’s train set Df to train some probe model g has some negative impact on the resulting probe’s performance, then we should ex…
Figure 5
Figure 5. Figure 5: Probe performance by amount of noise in data labels. does not necessarily convey to the user the meaning of such pixels or how they were used by the model. All of this guesswork still has to be done by the user. Concept-based explainable AI [39, 28, 47, 29] emerged fro…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 64 canonical work pages

  1. [1]

    Adebayo, M

    J. Adebayo, M. Muelly, I. Liccardi, and B. Kim. Debugging Tests for Model Explanations. In NeurIPS’20, 2020

    work page 2020

  2. [2]

    Alain and Y

    G. Alain and Y . Bengio. Understanding intermediate layers using linear classifier probes. In ICLR’17 Workshop Track, 2017

    work page 2017

  3. [3]

    M. G. Augasta and T. Kathirvalavakumar. Reverse Engineering the Neural Networks for Rule Extraction in Classification Problems.Neural Process. Lett., 35(2):131–150, 2012

    work page 2012

  4. [4]

    Belinkov

    Y . Belinkov. Probing Classifiers: Promises, Shortcomings, and Ad- vances. Comput. Linguistics, 48(1):207–219, 2022

    work page 2022

  5. [5]

    Belinkov, N

    Y . Belinkov, N. Durrani, F. Dalvi, H. Sajjad, and J. R. Glass. What do Neural Machine Translation Models Learn about Morphology? In ACL’17, 2017

    work page 2017

  6. [6]

    E. Chu, D. Roy, and J. Andreas. Are Visual Explanations Useful? A Case Study in Model-in-the-Loop Prediction. abs/2007.12248, 2020

    arXiv 2007

  7. [7]

    Crabbé and M

    J. Crabbé and M. van der Schaar. Concept activation regions: A general- ized framework for concept-based explanations. In NeurIPS’22, 2022

    work page 2022

  8. [8]

    Explainable Abstract Trains Dataset

    M. de Sousa Ribeiro, L. Krippahl, and J. Leite. Explainable Abstract Trains Dataset. CoRR, abs/2012.12115, 2020

    work page Pith review arXiv 2012

Show all 68 references
  1. [9]

    de Sousa Ribeiro, A

    M. de Sousa Ribeiro, A. Leote, and J. Leite. CUB-200-2011 At- tribute Revision, 2025. URL https://codelab.fct.unl.pt/di/is/neurosym/ cub_attr_rev

    2011

  2. [10]

    de Sousa Ribeiro, A

    M. de Sousa Ribeiro, A. Leote, and J. Leite. On the Performance of Concept Probing: The Influence of the Data. In Procs. of ECAI’25 , 2025

    2025

  3. [11]

    de Sousa Ribeiro, A

    M. de Sousa Ribeiro, A. Leote, and J. Leite. Concept Probing: Where to Find Human-Defined Concepts. In Procs. of NeSy’25, 2025

    2025

  4. [12]

    de Sousa Ribeiro, A

    M. de Sousa Ribeiro, A. Leote, and J. Leite. GTSRB-Concepts: Additional German Traffic Sign Recognition Benchmark Labels and Ontology, 2025. URL https://codelab.fct.unl.pt/di/is/neurosym/gtsrb_ concepts

    2025

  5. [13]

    de Sousa Ribeiro and J

    M. de Sousa Ribeiro and J. Leite. Aligning Artificial Neural Networks and Ontologies towards Explainable AI. In AAAI’21, 2021

    2021

  6. [14]

    J. Deng, W. Dong, R. Socher, L. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR’09, 2009

    2009

  7. [15]

    Durrani, H

    N. Durrani, H. Sajjad, F. Dalvi, and Y . Belinkov. Analyzing Individual Neurons in Pre-trained Language Models. In EMNLP, 2020

    2020

  8. [16]

    Ferreira, M

    J. Ferreira, M. de Sousa Ribeiro, R. Gonçalves, and J. Leite. Looking Inside the Black-Box: Logic-based Explanations for Neural Networks. In KR’22, 2022

    2022

  9. [17]

    Ghorbani, J

    A. Ghorbani, J. Wexler, J. Y . Zou, and B. Kim. Towards Automatic Concept-based Explanations. In NeurIPS’19, 2019

    2019

  10. [18]

    Y . Gong, G. Liu, Y . Xue, R. Li, and L. Meng. A survey on dataset quality in machine learning. Inf. Softw. Technol., 162:107268, 2023

    2023

  11. [19]

    Gurnee and M

    W. Gurnee and M. Tegmark. Language Models Represent Space and Time. In ICLR’24, 2024

    2024

  12. [20]

    G. V . Horn, O. M. Aodha, Y . Song, Y . Cui, C. Sun, A. Shepard, H. Adam, P. Perona, and S. J. Belongie. The INaturalist Species Clas- sification and Detection Dataset. In CVPR 2018, 2018

    2018

  13. [21]

    V . A. C. Horta, I. Tiddi, S. Little, and A. Mileo. Extracting knowledge from Deep Neural Networks through graph analysis. Future Gener . Comput. Syst., 120:109–118, 2021

    2021

  14. [22]

    Hupkes, S

    D. Hupkes, S. Veldhoen, and W. H. Zuidema. Visualisation and ’Di- agnostic Classifiers’ Reveal How Recurrent and Recursive Neural Net- works Process Hierarchical Structure. J. Artif. Intell. Res., 61:907–926, 2018

    2018

  15. [23]

    Ivanovs, R

    M. Ivanovs, R. Kadikis, and K. Ozols. Perturbation-based methods for explaining deep neural networks: A survey.Pattern Recognit. Lett., 150: 228–234, 2021

    2021

  16. [24]

    G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T. Liu. LightGBM: A Highly Efficient Gradient Boosting Decision Tree. In NeurIPS’17, 2017

    2017

  17. [25]

    B. Kim, M. Wattenberg, J. Gilmer, C. J. Cai, J. Wexler, F. B. Viégas, and R. Sayres. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCA V). InICML’18, 2018

    2018

  18. [26]

    P. W. Koh, T. Nguyen, Y . S. Tang, S. Mussmann, E. Pierson, B. Kim, and P. Liang. Concept Bottleneck Models. In Procs. of ICML’20, 2020

    2020

  19. [27]

    Krizhevsky

    A. Krizhevsky. Learning Multiple Layers of Features from Tiny Images. 2009

    2009

  20. [28]

    J. H. Lee, S. Lanza, and S. Wermter. From neural activations to con- cepts: A survey on explaining concepts in neural networks. CoRR, abs/2310.11884, 2023

    2023 arXiv

  21. [29]

    J. H. Lee, G. Mikriukov, G. Schwalbe, S. Wermter, and D. Wolter. Concept-based explanations in computer vision: Where are we and where could we go? CoRR, abs/2409.13456, 2024

    2024 arXiv

  22. [30]

    Linzen, E

    T. Linzen, E. Dupoux, and Y . Goldberg. Assessing the Ability of LSTMs to Learn Syntax-Sensitive Dependencies. Trans. Assoc. Com- put. Linguistics, 4:521–535, 2016

    2016

  23. [31]

    N. F. Liu, M. Gardner, Y . Belinkov, M. E. Peters, and N. A. Smith. Lin- guistic Knowledge and Transferability of Contextual Representations. In NAACL-HLT’19, 2019

    2019

  24. [32]

    Z. Liu, P. Luo, S. Qiu, X. Wang, and X. Tang. Deepfashion: Power- ing robust clothes recognition and retrieval with rich annotations. In CVPR’16, 2016

    2016

  25. [33]

    Lovering and E

    C. Lovering and E. Pavlick. Unit Testing for Concepts in Neural Net- works. Trans. Assoc. Comput. Linguistics, 10:1193–1208, 2022

    2022

  26. [34]

    Mu and J

    J. Mu and J. Andreas. Compositional explanations of neurons. In NeurIPS’20, 2020

    2020

  27. [35]

    C. G. Northcutt, L. Jiang, and I. L. Chuang. Confident Learning: Es- timating Uncertainty in Dataset Labels. J. Artif. Intell. Res. , 70:1373– 1411, 2021

    2021

  28. [36]

    Pálsson and Y

    A. Pálsson and Y . Björnsson. Empirical Evaluation of Concept Probing for Game-Playing Agents. In ECAI’24, 2024

    2024

  29. [37]

    Pimentel, N

    T. Pimentel, N. Saphra, A. Williams, and R. Cotterell. Pareto Probing: Trading Off Accuracy for Complexity. In EMNLP’20, 2020

    2020

  30. [38]

    Pimentel, J

    T. Pimentel, J. Valvoda, R. H. Maudslay, R. Zmigrod, A. Williams, and R. Cotterell. Information-Theoretic Probing for Linguistic Structure. In ACL’20, 2020

    2020

  31. [39]

    Poeta, G

    E. Poeta, G. Ciravegna, E. Pastor, T. Cerquitelli, and E. Baralis. Concept-based explainable artificial intelligence: A survey. CoRR, abs/2312.12936, 2023

    2023

  32. [40]

    Rebuffi, R

    S. Rebuffi, R. Fong, X. Ji, and A. Vedaldi. There and Back Again: Revisiting Backpropagation Saliency Methods. In CVPR, 2020

    2020

  33. [41]

    Why Should I Trust You?

    M. T. Ribeiro, S. Singh, and C. Guestrin. "Why Should I Trust You?": Explaining the Predictions of Any Classifier. In SIGKDD’16, 2016

    2016

  34. [42]

    Russakovsky and L

    O. Russakovsky and L. Fei-Fei. Attribute learning in large-scale datasets. In ECCV’10 Trends and Topics in Computer Vision Work- shops, 2010

    2010

  35. [43]

    Sandler, A

    M. Sandler, A. G. Howard, M. Zhu, A. Zhmoginov, and L. Chen. Mo- bileNetV2: Inverted Residuals and Linear Bottlenecks. In CVPR’18, 2018

    2018

  36. [44]

    Sanh and A

    V . Sanh and A. M. Rush. Low-Complexity Probing via Finding Subnet- works. In NAACL-HLT’21, 2021

    2021

  37. [45]

    Schmarje, V

    L. Schmarje, V . Grossmann, C. Zelenka, S. Dippel, R. Kiko, M. Oszust, M. Pastell, J. Stracke, A. Valros, N. V olkmann, and R. Koch. Is one annotation enough? - A data-centric image classification benchmark for noisy and ambiguous label estimation. In NeurIPS’22, 2022

    2022

  38. [46]

    G. P. J. Schmitz, C. Aldrich, and F. S. Gouws. ANN-DT: an algorithm for extraction of decision trees from artificial neural networks. IEEE Trans. Neural Networks, 10(6):1392–1401, 1999

    1999

  39. [47]

    Schwalbe and B

    G. Schwalbe and B. Finzel. A comprehensive taxonomy for explainable artificial intelligence: a systematic survey of surveys on methods and concepts. Data Min. Knowl. Discov., 38(5):3043–3101, 2024

    2024

  40. [48]

    Shen and T

    H. Shen and T. K. Huang. How Useful Are the Machine-Generated Interpretations to General Users? A Human Evaluation on Guessing the Incorrectly Predicted Labels. In HCOMP’20, 2020

    2020

  41. [49]

    Simonyan and A

    K. Simonyan and A. Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. In ICLR’15, 2015

    2015

  42. [50]

    H. Song, M. Kim, and J. Lee. Toward robustness in multi-label classi- fication: A data augmentation strategy against imbalance and noise. In AAAI’24, 2024

    2024

  43. [51]

    Stallkamp, M

    J. Stallkamp, M. Schlipsing, J. Salmen, and C. Igel. The German Traffic Sign Recognition Benchmark: A multi-class classification competition. In IJCNN’11, 2011

    2011

  44. [52]

    Sundararajan, A

    M. Sundararajan, A. Taly, and Q. Yan. Axiomatic Attribution for Deep Networks. In ICML’17, 2017

    2017

  45. [53]

    M. R. Taesiri, G. Nguyen, and A. Nguyen. Visual correspondence-based explanations improve AI robustness and human-AI team accuracy. In NeurIPS’22, 2022

    2022

  46. [54]

    Tenney, P

    I. Tenney, P. Xia, B. Chen, A. Wang, A. Poliak, R. T. McCoy, N. Kim, B. V . Durme, S. R. Bowman, D. Das, and E. Pavlick. What do you learn from context? probing for sentence structure in contextualized word representations. In ICLR’19, 2019

    2019

  47. [55]

    Tucker, P

    M. Tucker, P. Qian, and R. Levy. What if This Modified That? Syntac- tic Interventions with Counterfactual Embeddings. In ACL/IJCNLP’21, 2021

    2021

  48. [56]

    Vienna Convention on Road Traffic

    United Nations. Vienna Convention on Road Traffic. United Nations Publications, 1968

    1968

  49. [57]

    Villalobos, J

    P. Villalobos, J. Sevilla, T. Besiroglu, L. Heim, A. Ho, and M. Hobb- hahn. Machine Learning Model Sizes and the Parameter Gap. CoRR, abs/2207.02852, 2022

    2022 arXiv

  50. [58]

    C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR- 2011-001, 2011

    2011

  51. [59]

    Wang and S

    Y . Wang and S. H. Chung. Artificial intelligence in safety-critical sys- tems: a systematic review. Ind. Manag. Data Syst. , 122(2):442–470, 2022

    2022

  52. [60]

    Whatbird: Field guide to birds of north america

    WhatBird. Whatbird: Field guide to birds of north america. http://www. whatbird.com, 2022

    2022

  53. [61]

    B. Wu, C. Xu, X. Dai, A. Wan, P. Zhang, M. Tomizuka, K. Keutzer, and P. Vajda. Visual transformers: Token-based image representation and processing for computer vision. abs/2006.03677, 2020

    2006 arXiv

  54. [62]

    Yu and S

    H. Yu and S. Winkler. Image complexity and spatial information. In QoMEX’12, 2013

    2013

  55. [63]

    X. Zhao, Y . Yang, F. Zhou, X. Tan, Y . Yuan, Y . Bao, and Y . Wu. Rec- ognizing Part Attributes With Insufficient Data. In ICCV, 2019

    2019

  56. [64]

    B. Zhou, À. Lapedriza, A. Khosla, A. Oliva, and A. Torralba. Places: A 10 million image database for scene recognition. IEEE Trans. Pattern Anal. Mach. Intell., 40(6):1452–1464, 2018

    2018

  57. [65]

    Zhou and V

    Y . Zhou and V . Srikumar. DirectProbe: Studying Representations with- out Classifiers. In NAACL-HLT’21, 2021. A Additional Caltech-UCSD Birds-200-2011 Labels In this Appendix, we describe the procedure used to relabel a subset of the Caltech-UCSD Birds-200-2011 dataset (CUB) ...

    2021

  58. [66]

    Split the data into train and test images, according to the splits defined in the dataset

  59. [67]

    (b) Define a constant n = min({|P |, |N |, 500}) (c) Randomly select n samples from P : i

    For each of the train and test sets: (a) Split the data into a positive set ( P ), where the attribute is present, and a negative set ( N), where the attribute is absent, of samples, according to the original labels in the dataset. (b) Define a constant n = min({|P |, |N |, 50...

  60. [68]

    Model B, 2a is octagonal with a red ground bearing the word

    regarding additional concepts of interest defined in the 1968 Convention on Road Signs and Signals [56], which describes each type of traffic sign in the dataset based on visual features. These labels and ontology are made available in [12]. B.1 Ontology Development This ontol...

    1968

Pith tools

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