Pith. sign in

REVIEW 3 major objections 4 minor 60 references

FastCAV: Efficient Computation of Concept Activation Vectors for Explaining Deep Neural Networks

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read FastCAV computes a concept activation vector as the normalized difference between the concept-activation mean and the global mean in layer space, matching SVM-based CAVs while running up to 63.6x faster.

desk verdict FastCAV is a useful empirical speedup for CAV computation, but the SVM-equivalence theory is over-claimed and the drop-in framing needs a tail-risk qualifier. read the letter →

arxiv 2505.17883 v1 pith:DAWM2H3H submitted 2025-05-23 cs.LG cs.AIcs.CV

Laines Schmalwasser , Niklas Penzel , Joachim Denzler , Julia Niebling This is my paper
classification cs.LGcs.AIcs.CV
keywords conceptactivationvectorsTCAVexplainabilitylineardiscriminantanalysissupportvectormachinesneuralnetworkinterpretabilityfeaturesuperpositionconcept-basedexplanations
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

FastCAV is a method for computing Concept Activation Vectors (CAVs) — directions in a neural network's activation space that represent human-understandable concepts — without training a classifier. Instead of fitting a linear SVM, it takes the unit vector pointing from the mean of concept and random activations toward the mean of the concept activations, a single mean-difference computation. The paper argues the move is not a heuristic: under Gaussian, isotropic-covariance assumptions the direction is the Fisher LDA solution, and at high dimensionality the linear SVM normal vector converges to the same direction. Empirically, FastCAV is up to 63.6x faster (46.4x on average) across eight architectures, produces CAVs of comparable or better accuracy, and yields equivalent TCAV and ACE insights. If the claim holds, concept-based explanation stops being a bottleneck and can be run at the scale of per-layer, per-epoch audits of deep models.

What carries the argument

The load-bearing object is the normalized vector $$v^l_c \propto \frac{1}{|D_c|}\sum_{x\in D_c}\bigl(g_l(x)-\hat\mu_{D_c\cup D_r}\bigr),$$ the direction from the pooled activation mean to the concept-example mean, with the linear boundary completed by $b=-v^l_c\cdot\hat\mu_{D_c\cup D_r}$. It replaces the SVM objective with a single pass over activations. The identity that carries the argument is the reduction of Fisher discriminant analysis to this mean difference under isotropic within-class covariance, together with the known result that a linear SVM is a sparsified Fisher discriminant whose solution converges to the same vector in the $d \gg n$ regime. This chain is what turns a mean computation into a justified CAV rather than an arbitrary probe.

What would settle it

Pick a modern ViT or ConvNeXt layer and compute the eigenvalue spread (condition number) of the pooled within-class covariance for concept versus random activations, then compare FastCAV's direction with an SVM-trained CAV on the same data. If layers with large condition numbers consistently show cosine similarity far below the reported 0.8-0.9 range, or the fraction of concepts with accuracy gaps above 25 points grows well beyond the reported 2.8%, the drop-in equivalence claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a concept activation vector can be computed as the normalized mean-difference direction, $v^l_c \propto \frac{1}{|D_c|}\sum_{x\in D_c}(g_l(x)-\hat\mu_{D_c\cup D_r})$, where $D_c$ is the concept image set, $D_r$ the random set, and $\hat\mu$ the pooled activation mean. Under equal Gaussian classes the expected direction is proportional to $\mu_c-\mu_r$; with additional isotropic within-class covariance it equals the Fisher LDA normal vector, and via the known SVM-Fisher relationship the SVM-based CAV of prior work converges to the same vector when activation dimension greatly exceeds sample count. Across ResNet50, Inception-v3, ViT-B/16, ConvNeXt-XXLarge, ViT-H/14-CLIP, EVA-G/14, EVA-02-L/14 and RegNetY, the paper reports speeds 46.4x faster on average (up to 63.6x), CAV accuracies similar to or better than SVM-CAV, higher stability under resampled random sets, and matching TCAV and ACE conclusions. The paper also states the Gaussian-isotropy assumptions are strong and likely violated in practice, and reports that in 2.8% of CAVs the accuracy difference between the two methods exceeds 25 percentage points, with SVM favored in 1% of all cases.

Load-bearing premise

The concept and random activations are Gaussian, equally mixed, and have isotropic within-class covariance, so the within-class covariance matrix is a scalar multiple of the identity and the Fisher discriminant normal vector collapses to the mean difference; the paper itself calls these assumptions strong and likely false in practice.

Editorial extensions

If this is right

  • Concept-based explanations become affordable at per-layer and per-epoch granularity; the paper demonstrates this by tracking CAVs through all 90 epochs of a ResNet50 trained on ImageNet, a computation previously impractical.
  • TCAV and ACE can substitute FastCAV for their SVM component without changing pipelines: reported TCAV scores show the same qualitative conclusions, with smaller variance across resampled random sets.
  • Higher intra-method robustness means fewer resampled random sets are needed to get stable TCAV significance, reducing the cost of statistical testing further.
  • The theoretical equivalence establishes that in high-dimensional activation spaces, linear SVM training for CAVs is effectively solving for a mean difference, so the speedup does not come at the cost of a different semantic direction in the typical regime.

Reading between the lines

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

  • A direct extension the paper leaves implicit: since FastCAV's only source of variation across runs is the estimated global mean, the run-to-run variance in TCAV scores is likely dominated by SVM optimizer noise; replacing the SVM with FastCAV and recomputing TCAV confidence intervals would test this directly.
  • The mean-difference construction could be generalized to class-conditional concepts (replacing the global mean with a class-conditional mean) to attribute concept directions within specific classes, a variant the paper does not explore.
  • The reported 2.8% of large-accuracy-divergence cases suggests a practical diagnostic: estimate the eigenvalue spread of the within-class covariance for each concept and fall back to SVM training when it is high, making FastCAV a front-end filter rather than a full replacement.
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 / 4 minor

Summary. FastCAV replaces SVM-based CAV training with a one-pass mean computation: the normalized difference between the concept activation mean and the global activation mean. The paper claims speedups of up to 63.6x (46.4x on average) with comparable CAV accuracy and similarity across eight ImageNet architectures, a medical chest X-ray task, and downstream TCAV and ACE analyses, and demonstrates the method's usefulness by tracking concept evolution during ResNet50 training. A theoretical section argues that FastCAV is equivalent to Fisher LDA under Gaussian isotropic within-class covariance and, via Shashua's support-vector sparsification result, to linear SVM when d >> n.

Significance. The empirical contribution is substantial and carefully executed: the evaluation spans eight architectures, multiple layers, repeated random sets, multiple concepts, several linear classifiers, and downstream explanation methods, with consistent results. The speed advantage of a simple mean arithmetic estimator over iterative SVM training is robust and the training-time tracking experiment is a genuine new capability. If the theoretical equivalence claim is either proven under precise conditions or honestly weakened to an approximate/empirical statement, and the drop-in claim is qualified by the acknowledged failure cases, the paper would be a useful and practical contribution to concept-based interpretability. The present version, however, overstates the theoretical support for the method.

major comments (3)
  1. [§3.3] The claimed exact equivalence to linear SVM is not established. Shashua (1999) equates the linear SVM solution with the Fisher discriminant computed on the support-vector set, not on the full data. Appendix A.2 (Table 3) reports support-vector fractions between 52.31% and 99.67%; whenever this fraction is below 100%, the support-vector set is selected by the margin and its within-class scatter is not the full-data scatter, so the SVM direction need not coincide with the full-data LDA direction even under isotropic Gaussian class-conditional distributions. The heuristic that "d >> n implies many support vectors" plus the empirical percentages does not bridge this gap. In addition, Eq. (6) states that the expectation of the normalized CAV is proportional to the mean difference, but expectation and normalization do not commute; this needs a concentration argument or an explicit statement that the proportionality is only approximate. Please either provide a precise theorem with conditions under which the two directions are equal (or a bound on their angle), or clearly label the equivalence as approximate and empirically motivated.
  2. [Appendix B.2.2] The tail behavior reported by the authors contradicts the unqualified "drop-in replacement" claim made in the abstract and conclusion. Appendix B.2.2 states that 2.8% of all computed CAVs have accuracy differences above 25 percentage points, with SVM winning in 1% of cases, and that one ViT-B/16 encoder layer shows a 40-point gap (95% vs 55%). These are precisely the cases where the isotropic Gaussian assumption fails, and Table 1 reports only averages, so it does not quantify the risk faced by a user applying FastCAV to a new concept or layer. Please add a clearly visible caveat to the abstract and conclusion, and ideally provide a practical diagnostic (e.g., checking separability or comparing against a small SVM holdout) to detect such cases.
  3. [§3.3, Eq. (7)] Equation (7) contains a visible typo in the definition of the within-class covariance estimator: the second sum uses (gl(x) − μ̂_r)(gl(x) − μ̂_c)⊤, where the second factor should be (gl(x) − μ̂_r)⊤. Because the equation is the basis for the LDA reduction that motivates FastCAV, this should be corrected, and the surrounding derivation should be checked for similar indexing errors.
minor comments (4)
  1. [Appendix A.3, Figure 10] The caption "It can be seen that our method results in similar results" is uninformative and appears to be a leftover; it should describe the runtime scaling with n and d and label the plotted curves clearly.
  2. [Appendix A.2, Table 3] The header "∅ Activation Dimensionality" uses the empty-set symbol where "Mean" or "Average" is intended; please use conventional notation.
  3. [Figures 3, 11–14] The method label "FastCav" is inconsistent with the "FastCAV" spelling used in the text; please unify the spelling across all figures and captions.
  4. [Appendix A.2] The support-vector percentages are reported only for pre-final layers; given the theoretical argument, reporting the per-layer spread of support-vector fractions would be informative, since the argument's validity depends on the fraction being close to 1.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: FastCAV is a closed-form estimator benchmarked against an external SVM baseline; its equivalence arguments rest on external classical results, not on the paper's own output.

full rationale

FastCAV is defined directly by Eq. (3) as the normalized mean difference between concept activations and the global mean, with no parameter fitted to any target that is later called a prediction. The theoretical derivation in Section 3.3 uses standard Gaussian-mixture calculations (Eqs. 5-6) and the classical fact that under isotropic within-class covariance the Fisher LDA normal direction is proportional to the mean difference; the identification with linear SVM relies on the externally published equivalence of Shashua (1999) and on externally published support-vector proliferation results (Muthukumar et al., 2021; Hsu et al., 2021), not on results of the present paper. The empirical comparisons in Sections 4.1-4.3 benchmark FastCAV against an independently computed SVM baseline on held-out validation data (Appendix B.1), so the reported accuracy and similarity measurements are external checks rather than consequences of the method's construction. The acknowledged 2.8% large-accuracy-difference cases (Appendix B.2.2) are an empirical limitation of the stated Gaussian-isotropy assumptions, but no fitted constant is fed back into the method; this is a correctness and robustness caveat, not circularity. The only self-citation (Penzel et al., 2022) appears as a supporting aside on training dynamics in Appendix B.5 and is not load-bearing, and the typo in Eq. (7) does not make the derivation circular. Overall, the derivation chain is self-contained and the central claim is tested against an external baseline, so there is no significant circularity.

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

The method itself has no fitted parameters, but the theoretical equivalence claim rests on three assumptions: Gaussian activations, isotropic covariance, and an overparameterized regime with near-full support vector sets. The isotropic covariance assumption is the most fragile because it is contradicted by typical neural network activation structure, and the paper explicitly acknowledges that these assumptions are strong (Appendix B.2.2).

assumptions (3)
  • domain assumption Activations for concept and random images follow multivariate Gaussian distributions with equal mixture weights.
    Appendix A.1 assumes gl(x) ~ N(mu_c, Sigma_c) and gl(x) ~ N(mu_r, Sigma_r) to derive E[global mean] = (mu_c + mu_r)/2. This is a strong distributional assumption about deep network activations.
  • ad hoc to paper Within-class covariance is isotropic, i.e., Sigma^{-1} is proportional to the identity matrix.
    Section 3.3, after Eq. 7, states that the Fisher discriminant solution reduces to FastCAV only under this explicit additional assumption. Deep network activations generally have highly non-isotropic covariance, so this assumption is introduced to make the derivation work.
  • domain assumption The dimensionality is much larger than the sample size, and the support vector fraction is close to 1.
    Section 3.3 cites Muthukumar et al. and Hsu et al. and Table 3 empirically shows high support vector ratios. This is used to connect Shashua's SVM-as-sparsified-LDA result to the full-sample LDA solution, but the connection is not a formal proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FastCAV: Efficient Computation of Concept Activation Vectors for Explaining Deep Neural Networks." pith.science (2026). https://pith.science/paper/DAWM2H3H

@misc{pith2026250517883,
  author       = {Pith},
  title        = {Pith review of: FastCAV: Efficient Computation of Concept Activation Vectors for Explaining Deep Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DAWM2H3H}},
  note         = {Machine review of arXiv:2505.17883}
}
read the original abstract

Concepts such as objects, patterns, and shapes are how humans understand the world. Building on this intuition, concept-based explainability methods aim to study representations learned by deep neural networks in relation to human-understandable concepts. Here, Concept Activation Vectors (CAVs) are an important tool and can identify whether a model learned a concept or not. However, the computational cost and time requirements of existing CAV computation pose a significant challenge, particularly in large-scale, high-dimensional architectures. To address this limitation, we introduce FastCAV, a novel approach that accelerates the extraction of CAVs by up to 63.6x (on average 46.4x). We provide a theoretical foundation for our approach and give concrete assumptions under which it is equivalent to established SVM-based methods. Our empirical results demonstrate that CAVs calculated with FastCAV maintain similar performance while being more efficient and stable. In downstream applications, i.e., concept-based explanation methods, we show that FastCAV can act as a replacement leading to equivalent insights. Hence, our approach enables previously infeasible investigations of deep models, which we demonstrate by tracking the evolution of concepts during model training.

Figures

Figures reproduced from arXiv: 2505.17883 by the authors.

Figure 1
Figure 1. Comparison of computational efficiency between [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Schematic illustration of FastCAV in two dimen [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. TCAV scores for various GoogleNet (Szegedy et al., 2015) layers. We compare the concepts “polka-dotted”, “striped”, and “zigzagged” for the class ladybug using FastCAV against SVM-CAV. We follow (Kim et al., 2018) and mark CAVs that are not statistically significant with “*”. Testing with Concept Activation Vectors (TCAV) TCAV calculates the directional derivative of a selected class k with respect to the activation… view at source ↗
Figures from the paper (17 more)
Figure 4
Figure 4. Figure 4: Most salient concepts discovered by ACE ( [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Sensitivity analysis of FastCAV to (a) the number [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Evolution of various CAVs during training of a ResNet50 ( [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Evolution of various CAVs during training of a [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: It can be seen that our method results in similar results. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Comparison of computational efficiency between FastCAV and the established SVM-CAV for CAVs across [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Left: Histogram of accuracy differences between FastCAV and SVM-CAV. Note the logarithmic y-axis scale. [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Example of the concepts “red”, “yellow”, “blue” and “green” for class [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Example of the concepts “red”, “yellow”, “blue” and “green” for class [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 13
Figure 13. Figure 13: Example of the concepts “red”, “yellow”, “blue” and “green” for class [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Example of the concepts “polka-dotted”, “striped” and “zigzagged” for class [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 15
Figure 15. Figure 15: Comparison of the most salient concepts discovered by ACE ( [PITH_FULL_IMAGE:figures/full_fig_p022_15.png]
Figure 16
Figure 16. Figure 16: Comparison of the most salient concepts discovered by ACE ( [PITH_FULL_IMAGE:figures/full_fig_p023_16.png]
Figure 17
Figure 17. Figure 17: Comparison of the most salient concepts discovered by ACE ( [PITH_FULL_IMAGE:figures/full_fig_p024_17.png]
Figure 18
Figure 18. Figure 18: Additional visualizations pertaining to the lefthand side in Figure [PITH_FULL_IMAGE:figures/full_fig_p025_18.png]
Figure 19
Figure 19. Figure 19: Additional visualizations for the middle plot in Figure [PITH_FULL_IMAGE:figures/full_fig_p026_19.png]
Figure 20
Figure 20. Figure 20: Percentage of learned concepts by CAVs and block during training of a ResNet50 ( [PITH_FULL_IMAGE:figures/full_fig_p027_20.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 32 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Understanding intermediate layers using linear classifier probes

    Alain, G. Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644, 2016

    arXiv 2016

  3. [3]

    Perceptual symbol systems

    Barsalou, L. Perceptual symbol systems. The Behavioral and brain sciences/Cambridge University Press, 1999

    work page 1999

  4. [4]

    Network dissection: Quantifying interpretability of deep visual representations

    Bau, D., Zhou, B., Khosla, A., Oliva, A., and Torralba, A. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 6541--6549, 2017

    work page 2017

  5. [5]

    Understanding the role of individual units in a deep neural network

    Bau, D., Zhu, J.-Y., Strobelt, H., Lapedriza, A., Zhou, B., and Torralba, A. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 117 0 (48): 0 30071--30078, 2020

    work page 2020

  6. [6]

    Bishop, C. M. Pattern recognition and machine learning, volume 4. Springer, 2006

    work page 2006

  7. [7]

    L., Anil, C., Denison, C., Askell, A., et al

    Bricken, T., Templeton, A., Batson, J., Chen, B., Jermyn, A., Conerly, T., Turner, N. L., Anil, C., Denison, C., Askell, A., et al. Towards Monosemanticity: Decomposing Language Models With Dictionary Learning. Anthropic, 2023

    work page 2023

  8. [8]

    and Lin, C.-J

    Chang, C.-C. and Lin, C.-J. Libsvm: A library for support vector machines. ACM transactions on intelligent systems and technology (TIST), 2 0 (3): 0 1--27, 2011

    work page 2011

Show all 60 references
  1. [9]

    Training a support vector machine in the primal

    Chapelle, O. Training a support vector machine in the primal. Neural Computation, 19 0 (5): 0 1155--1178, 2007. doi:10.1162/neco.2007.19.5.1155

    2007 doi

  2. [10]

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

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021

    2021

  3. [11]

    Toy models of superposition

    Elhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Kravec, S., Hatfield-Dodds, Z., Lasenby, R., Drain, D., Chen, C., et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022

    2022 arXiv

  4. [12]

    Liblinear: A library for large linear classification

    Fan, R.-E., Chang, K.-W., Hsieh, C.-J., Wang, X.-R., and Lin, C.-J. Liblinear: A library for large linear classification. Journal of machine learning research, 9 0 (Aug): 0 1871--1874, 2008

    2008

  5. [13]

    Eva: Exploring the limits of masked visual representation learning at scale

    Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., and Cao, Y. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 19358--19369, 2023

    2023

  6. [14]

    Eva-02: A visual representation for neon genesis

    Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., and Cao, Y. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149: 0 105171, 2024

    2024

  7. [15]

    Y., and Kim, B

    Ghorbani, A., Wexler, J., Zou, J. Y., and Kim, B. Towards automatic concept-based explanations. In Advances in Neural Information Processing Systems, pp.\ 9273--9282, 2019

    2019

  8. [16]

    Distilling blackbox to interpretable models for efficient transfer learning

    Ghosh, S., Yu, K., and Batmanghelich, K. Distilling blackbox to interpretable models for efficient transfer learning. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pp.\ 628--638. Springer, 2023

    2023

  9. [17]

    Decoding the thought vector, 2016

    Goh, G. Decoding the thought vector, 2016. URL https://gabgoh.github.io/ThoughtVectors/. Accessed: 2025-01-07

    2016

  10. [18]

    Regression concept vectors for bidirectional explanations in histopathology

    Graziani, M., Andrearczyk, V., and M \"u ller, H. Regression concept vectors for bidirectional explanations in histopathology. In Understanding and Interpreting Machine Learning in Medical Image Computing Applications: First International Workshops, MLCN 2018, DLF 2018, and iM...

    2018

  11. [19]

    Concept distillation: leveraging human-centered explanations for model improvement

    Gupta, A., Saini, S., and Narayanan, P. Concept distillation: leveraging human-centered explanations for model improvement. Advances in Neural Information Processing Systems, 36, 2024

    2024

  12. [20]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 770--778, 2016

    2016

  13. [21]

    On the proliferation of support vectors in high dimensions

    Hsu, D., Muthukumar, V., and Xu, J. On the proliferation of support vectors in high dimensions. In International Conference on Artificial Intelligence and Statistics, pp.\ 91--99. PMLR, 2021

    2021

  14. [22]

    Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger, K. Q. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 4700--4708, 2017

    2017

  15. [23]

    Lg-cav: Train any concept activation vector with language guidance

    Huang, Q., Song, J., Xue, M., Zhang, H., Hu, B., Wang, H., Jiang, H., Wang, X., and Song, M. Lg-cav: Train any concept activation vector with language guidance. arXiv preprint arXiv:2410.10308, 2024

    2024 arXiv

  16. [24]

    Timm leaderboard, 2025

    HuggingFace. Timm leaderboard, 2025. URL https://huggingface.co/spaces/timm/leaderboard. Accessed: [05/01/2025]

    2025

  17. [25]

    Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison

    Irvin, J., Rajpurkar, P., Ko, M., Yu, Y., Ciurea-Ilcus, S., Chute, C., Marklund, H., Haghgoo, B., Ball, R., Shpanskaya, K., et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Thirty-Third AAAI Conference on Artificial Intelligence, 2019

    2019

  18. [26]

    E., Pollard, T

    Johnson, A. E., Pollard, T. J., Greenbaum, N. R., Lungren, M. P., Deng, C.-y., Peng, Y., Lu, Z., Mark, R. G., Berkowitz, S. J., and Horng, S. Mimic-cxr-jpg, a large publicly available database of labeled chest radiographs. arXiv preprint arXiv:1901.07042, 2019

    1901 arXiv

  19. [27]

    Visualizing and understanding recurrent networks

    Karpathy, A., Johnson, J., and Fei-Fei, L. Visualizing and understanding recurrent networks. arXiv preprint arXiv:1506.02078, 2015

    2015 arXiv

  20. [28]

    Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav), 2018

    Kim, B., Wattenberg, M., Gilmer, J., Cai, C., Wexler, J., Viegas, F., and Sayres, R. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav), 2018. URL https://arxiv.org/abs/1711.11279

    2018 arXiv

  21. [29]

    A convnet for the 2020s

    Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., and Xie, S. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 11976--11986, 2022

    2022

  22. [30]

    and Hutter, F

    Loshchilov, I. and Hutter, F. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  23. [31]

    Text2concept: Concept activation vectors directly from text

    Moayeri, M., Rezaei, K., Sanjabi, M., and Feizi, S. Text2concept: Concept activation vectors directly from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 3744--3749, 2023

    2023

  24. [32]

    Classification vs regression in overparameterized regimes: Does the loss function matter? Journal of Machine Learning Research, 22 0 (222): 0 1--69, 2021

    Muthukumar, V., Narang, A., Subramanian, V., Belkin, M., Hsu, D., and Sahai, A. Classification vs regression in overparameterized regimes: Does the loss function matter? Journal of Machine Learning Research, 22 0 (222): 0 1--69, 2021

    2021

  25. [33]

    A., and Gal, Y

    Nicolson, A., Schut, L., Noble, J. A., and Gal, Y. Explaining explainability: Understanding concept activation vectors, 2024. URL https://arxiv.org/abs/2404.03713

    2024 arXiv

  26. [34]

    and Weng, T.-W

    Oikarinen, T. and Weng, T.-W. Clip-dissect: Automatic description of neuron representations in deep vision networks. arXiv preprint arXiv:2204.10965, 2022

    2022 arXiv

  27. [35]

    and Weng, T.-W

    Oikarinen, T. and Weng, T.-W. Linear explanations for individual neurons. arXiv preprint arXiv:2405.06855, 2024

    2024 arXiv

  28. [36]

    Feature visualization

    Olah, C., Mordvintsev, A., and Schubert, L. Feature visualization. Distill, 2 0 (11): 0 e7, 2017

    2017

  29. [37]

    Zoom in: An introduction to circuits

    Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M., and Carter, S. Zoom in: An introduction to circuits. Distill, 5 0 (3): 0 e00024--001, 2020

    2020

  30. [38]

    J., Wiegand, T., Samek, W., and Lapuschkin, S

    Pahde, F., Dreyer, M., Weckbecker, M., Weber, L., Anders, C. J., Wiegand, T., Samek, W., and Lapuschkin, S. Navigating neural space: Revisiting concept activation vectors to overcome directional divergence. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  31. [39]

    Pytorch: An imperative style, high-performance deep learning library

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019

    2019

  32. [40]

    Scikit-learn: Machine learning in python

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., et al. Scikit-learn: Machine learning in python. Journal of machine learning research, 12 0 (Oct): 0 2825--2830, 2011

    2011

  33. [41]

    Investigating neural network training on a feature level using conditional independence

    Penzel, N., Reimers, C., Bodesheim, P., and Denzler, J. Investigating neural network training on a feature level using conditional independence. In ECCV Workshop on Causality in Vision (ECCV-WS), pp.\ 383--399, Cham, 2022. Springer Nature Switzerland. ISBN 978-3-031-25075-0. d...

    2022 doi

  34. [42]

    T., Wei, J., Wei, M

    Pfau, J., Young, A. T., Wei, J., Wei, M. L., and Keiser, M. J. Robust semantic interpretability: Revisiting concept activation vectors. arXiv preprint arXiv:2104.02768, 2021

    2021 arXiv

  35. [43]

    Platt, J. C. Fast training of support vector machines using sequential minimal optimization. In Advances in Kernel Methods: Support Vector Learning. The MIT Press, 12 1998. ISBN 9780262283199. doi:10.7551/mitpress/1130.003.0016. URL https://doi.org/10.7551/mitpress/1130.003.0016

    1998 doi

  36. [44]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp.\ 8748--8763. PMLR, 2021

    2021

  37. [45]

    Imagenet large scale visual recognition challenge

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115: 0 211--252, 2015

    2015

  38. [46]

    Best of both worlds: local and global explanations with human-understandable concepts

    Schrouff, J., Baur, S., Hou, S., Mincu, D., Loreaux, E., Blanes, R., Wexler, J., Karthikesalingam, A., and Kim, B. Best of both worlds: local and global explanations with human-understandable concepts. arXiv preprint arXiv:2106.08641, 2021

    2021 arXiv

  39. [47]

    On the relationship between the support vector machine for classification and sparsified fisher's linear discriminant

    Shashua, A. On the relationship between the support vector machine for classification and sparsified fisher's linear discriminant. Neural Processing Letters, 9 0 (2): 0 129--139, Apr 1999. ISSN 1573-773X. doi:10.1023/A:1018677409366. URL https://doi.org/10.1023/A:1018677409366

    1999 doi

  40. [48]

    and Tishby, N

    Shwartz-Ziv, R. and Tishby, N. Opening the black box of deep neural networks via information. arXiv preprint arXiv:1703.00810, 2017

    2017 arXiv

  41. [49]

    Singh, M., Gustafson, L., Adcock, A., Reis, V. d. F., Gedik, B., Kosaraju, R. P., Mahajan, D., Girshick, R., Doll 'a r, P., and van der Maaten, L. Revisiting Weakly Supervised Pre-Training of Visual Perception Models . In CVPR, 2022

    2022

  42. [50]

    Using causal analysis for conceptual deep learning explanation

    Singla, S., Wallace, S., Triantafillou, S., and Batmanghelich, K. Using causal analysis for conceptual deep learning explanation. In Medical Image Computing and Computer Assisted Intervention--MICCAI 2021: 24th International Conference, Strasbourg, France, September 27--Octobe...

    2021

  43. [51]

    Intriguing properties of neural networks

    Szegedy, C. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013

    2013 arXiv

  44. [52]

    Going deeper with convolutions

    Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich, A. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 1--9, 2015

    2015

  45. [53]

    Rethinking the inception architecture for computer vision

    Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wojna, Z. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 2818--2826, 2016

    2016

  46. [54]

    Scaling monosemanticity: Extracting interpretable features from claude 3 sonnet

    Templeton, A. Scaling monosemanticity: Extracting interpretable features from claude 3 sonnet. Anthropic, 2024

    2024

  47. [55]

    Statistical learning theory

    Vapnik, V. Statistical learning theory. John Wiley & Sons, 2: 0 831--842, 1998

    1998

  48. [56]

    Pytorch image models

    Wightman, R. Pytorch image models. https://github.com/rwightman/pytorch-image-models, 2019

    2019

  49. [57]

    L., Gugger, S., Drame, M., Lhoest, Q., and Rush, A

    Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., Davison, J., Shleifer, S., von Platen, P., Ma, C., Jernite, Y., Plu, J., Xu, C., Scao, T. L., Gugger, S., Drame, M., Lhoest, Q., and Rush, A. M. Huggingface's tr...

    1910 arXiv

  50. [58]

    On completeness-aware concept-based explanations in deep neural networks

    Yeh, C.-K., Kim, B., Arik, S., Li, C.-L., Pfister, T., and Ravikumar, P. On completeness-aware concept-based explanations in deep neural networks. Advances in neural information processing systems, 33: 0 20554--20565, 2020

    2020

  51. [59]

    A., Shechtman, E., and Wang, O

    Zhang, R., Isola, P., Efros, A. A., Shechtman, E., and Wang, O. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 586--595, 2018

    2018

  52. [60]

    A., and Rubinstein, B

    Zhang, R., Madumal, P., Miller, T., Ehinger, K. A., and Rubinstein, B. I. Invertible concept-based explanations for cnn models with non-negative concept activation vectors. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pp.\ 11682--11690, 2021

    2021

Pith tools

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