Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Guaranteed Optimal Compositional Explanations for Neurons

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read This paper claims a best-first search that provably returns the optimal logical explanation of a neuron's spatial alignment.

desk verdict A genuinely new first algorithm for guaranteed-optimal compositional explanations; the central admissibility gap for mixed-operator paths is real but likely closable, and the '10-40% suboptimal' headline overstates because it includes same-IoU alternatives. read the letter →

arxiv 2511.20934 v2 pith:KHSG4KOL submitted 2025-11-25 cs.AI cs.CVcs.LG

Biagio La Rosa , Leilani H. Gilpin This is my paper · ORCID
classification cs.AIcs.CVcs.LG
keywords compositionalexplanationsneuroninterpretabilitybest-firstsearchheuristicIntersectionoverUnionoptimalityguaranteebeamconceptalignment
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

This paper sets out to establish that the best logical explanation of what a neuron detects—the combination of concepts whose locations overlap its activation pattern most—can be computed exactly, not merely approximated. The route is a decomposition of the overlap score into quantities that can be estimated cheaply, followed by a best-first search that expands partial formulas in order of their estimated ceiling and prunes using an estimated floor. The authors prove that, under the field's standard assumptions (distinct concepts, incremental left-associative formulas, 00-preserving logical operators), the first complete formula returned is the one with the highest true alignment. If the proof holds, earlier beam-search explanations were suboptimal in 10–40% of cases involving overlapping concepts, and the gap can be closed in roughly the time a plain beam search took.

What carries the argument

The central object is the decomposed Intersection-over-Union (dIoU), which splits the alignment calculation into four per-sample counts—unique intersections, common intersections, unique extras, common extras—so that the effect of each logical operator (OR, AND, AND NOT) on each count is predictable. The matching heuristic builds Top and Bottom vectors from the per-concept values of these counts to estimate, for any partial formula, the maximum and minimum IoU achievable by any continuation of bounded length. The algorithm that carries the argument is best-first search with an admissible heuristic: it prunes nodes whose estimated maximum is below the current best minimum, and it is this comb

What would settle it

On a small synthetic dataset with two overlapping concepts and a single neuron, exhaustively enumerate every legal formula of length at most three and compute the true IoU for each; then compute the heuristic's estimated maximum for every partial formula and check whether any estimate is below the true best IoU of a formula extending it. A single counterexample would show the heuristic is inadmissible and the optimality proof does not cover all paths.

Watch

Extended reading notes

Core claim

The central claim is that guaranteed optimal compositional explanations are feasible. The paper introduces the decomposed Intersection-over-Union (dIoU) score, which separates locations into unique and common elements and separates neuron-active locations into intersections and extras, and proves that dIoU equals the standard IoU whenever the logical operators are 00-preserving. On top of this decomposition it builds a heuristic that, for any partial formula, upper-bounds the best IoU achievable by extending it and lower-bounds the worst, using precomputed Top and Bottom vectors of individual concept contributions. The optimal algorithm is a best-first search that visits nodes in descending

Load-bearing premise

The guarantee assumes that the path heuristic never overestimates the best alignment reachable from any partial explanation, and that property is formally proven for paths using one operator but only sketched for paths that mix multiple operators.

Editorial extensions

If this is right

  • If the central claim is correct, any existing beam-search explanation can be replaced by a provably optimal one without leaving the standard assumptions of the field, at a comparable computational cost.
  • The 10–40% discrepancy measured in overlapping-concept settings means that a meaningful fraction of previously reported compositional explanations were not the best-aligned ones; this can now be detected and corrected.
  • A beam-search variant guided by the same heuristic matches or beats the runtime of prior guided beam search while finding the same explanations as exhaustive search, removing the need to keep annotations in memory.
  • The heuristic's independence from spatial information makes the optimal-search framework usable in non-vision settings where concept locations are defined differently, such as audio or text.

Reading between the lines

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

  • If the mixed-operator path estimates are later shown to be inadmissible in some corner of the state space, the guarantee would degrade gracefully to single-operator explanations; the empirical suboptimality of beam search would still hold but without a formal optimality certificate.
  • The same decomposition could be repurposed for other set-overlap metrics (for instance, Dice or other Jaccard-style scores) by re-deriving the Top and Bottom vectors, provided the operators remain 00-preserving.
  • A practical audit pipeline could run the beam variant first to flag units whose explanations are likely suboptimal, then invoke the optimal algorithm only for those units, making the guarantee affordable at large scale.
  • The 10–40% suboptimality rate suggests that interpretability conclusions drawn from beam-search explanations of overlapping-concept datasets may be systematically imprecise; re-running standard benchmarks with the optimal algorithm could yield a corrected catalogue of neuron explanations.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a framework for computing guaranteed optimal compositional explanations of neurons. It introduces a decomposition of the IoU objective (dIoU) into unique/common intersection and extra sets, designs a path heuristic that estimates the best continuation of any partial label, and presents a best-first search algorithm intended to return the globally optimal explanation within the state space defined by distinct concepts, left-associative formulas, and 00-preserving operators. The paper also proposes a beam-search variant guided by the heuristic and reports experiments on Cityscapes, ADE20K, and Broden assessing runtime and the suboptimality of prior beam-search explanations.

Significance. If the optimality claim is made fully rigorous, this would be a valuable contribution: the dIoU identity (Lemma 1) is exact, no fitted parameters enter the proof, and the work directly addresses a gap in the compositional-explanations literature. The proposed heuristic and beam variant are practically relevant and the empirical finding that many beam-search explanations are suboptimal is important. However, the current proof leaves a load-bearing gap for mixed-operator paths, and there are concrete errors in the definition of the disjointness matrix and in the reported empirical range.

major comments (3)
  1. [Section 2.3.2 / Appendix E.3 / F.3] The optimality guarantee requires the path heuristic to be admissible for the full state space of left-associative formulas with arbitrary operator sequences. Section 2.3.2 asserts that mixed-operator paths are handled by taking the max/min of exclusive-path estimates, but Appendix E.3 only analyzes the OR+AND pair (Eqs. 54-59) and Appendix F.3 proves admissibility for exclusive OR, AND, and AND-NOT paths. Algorithm 1 expands nodes with every allowed connective (line 49) and prunes nodes whose estimated dIoU_max falls below the global dIoU_min (lines 52-55), so an inadmissible mixed-path estimate could discard the optimum. The missing lemma is plausible (the 1-set of any formula built from L plus t concepts is a subset of the union of those concepts' positive supports), but it is not stated or proved. This is load-bearing for the central claim.
  2. [Section 2.3.1, Eq. (5)] The disjointness matrix is defined as D[k1,k2]=1 iff M_{k1}[x,j] != M_{k2}[x,j] for every location. That is the condition that the two concept matrices are logical complements, not the condition that the concepts have no annotation overlap. For typical disjoint concepts with locations where neither concept is annotated, D=0, so the exact 'disjoint' branch of Section 2.3.1 would not be entered. This contradicts the stated purpose of D and can materially affect the tightness of the estimates and the reported efficiency. The condition should be, e.g., D[k1,k2]=1 iff no location has both annotations.
  3. [Abstract / Section 3.2 / Table 2] The abstract and Section 3.2 claim that '10-40%' of beam-search explanations are suboptimal when overlapping concepts are involved. Table 2 reports Diff = 8% for ResNet, 22% for AlexNet, and 39% for DenseNet. The lower endpoint 10% is contradicted by the 8% ResNet row. If 'Diff' means something different from the suboptimality rate, the text must say so; as written, the stated range is not supported by the table.
minor comments (5)
  1. [Section 2.3.1, footnote 1] The footnote says 'cat AND NOT dog is always true.' This is not a valid logical statement by itself; presumably the intended meaning is that when the two concepts are disjoint, 'cat AND NOT dog' is equivalent to 'cat.' Please reword.
  2. [Appendix F.1, Eq. (62)] The displayed constraint uses |EU_min|+|EC_max| in the denominator, but the surrounding proof and requirement (64) concern EC_min. This looks like a typo and should be corrected for readability.
  3. [Section 3.1] The statement that 'both MMESH and our beam variant converge to the same solutions as those found by the extensive search (Vanilla Beam), since they use admissible heuristics' is not justified: an admissible heuristic alone does not guarantee that a bounded beam search reproduces the full beam search's outcome. If this equality was verified empirically, say so; otherwise remove or qualify the claim.
  4. [Section 3.2] Category 3 explanations are described as if the optimal algorithm 'correctly identifies' the alignment even though the table says these explanations have the same IoU. Under the formal objective Eq. (1), both formulas are equally optimal; semantic/visual preferences should not be presented as an optimality defect of beam search.
  5. [Section 2.1] The formula for the number of combinations is garbled in the typesetting: 'Pn k=1 nk−1 o Q k(|L1|−k)' should be a readable product/summation. Also, Table 4 has formatting issues in the mean±std columns.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dIoU decomposition and optimality proof are self-contained; the main weakness is an unproved mixed-path admissibility lemma, which is a proof gap rather than circularity.

full rationale

The paper's central derivation is self-contained rather than circular. Lemma 1 (Appendix A) proves that the proposed dIoU score equals the standard IoU whenever operators are 00-preserving, so the decomposition is a proven identity, not a definition that assumes the answer. The heuristic bounds in Section 2.3 and Appendix E are derived from set-inclusion inequalities and from Top/Bottom cumulative sums over the dataset; these are data-derived upper/lower bounds on any future concept addition, not parameters fitted to the optimal explanation or to the IoU values being predicted. Appendix F proves admissibility of the exclusive OR/AND/AND NOT path heuristics and uses those bounds plus exhaustive expansion/pruning to establish optimality; the proof does not invoke any result from the authors' prior work. The only use of self-citations ([20], [21]) is as background and as a baseline for comparison; the MMESH paper is not used to justify the optimality claim. The most significant weakness is that Section 2.3.2/E.3 asserts, rather than proves, admissibility for mixed-operator paths by taking max/min over exclusive-path estimates; this is a proof gap in the optimality guarantee, but it is a completeness/correctness risk, not a circular reduction of the conclusion to its inputs.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The central claim rests on domain assumptions inherited from the compositional-explanation literature and on one unproved mixed-operator admissibility assertion. There are no fitted free parameters and no new physical or ontological entities; the decomposed quantities are partitions of existing annotation and activation locations.

assumptions (5)
  • domain assumption Assumption 1: concepts in an explanation are distinct.
    Section 2.1; restricts the search space and is inherited from prior compositional-explanation work.
  • domain assumption Assumption 2: formulas are built incrementally, left-associatively, with an atomic concept as the right operand at each step.
    Section 2.1; the optimality guarantee covers only this grammar, not arbitrary Boolean formula trees.
  • domain assumption Assumption 3: logical operators are 00-preserving (OR, AND, AND NOT).
    Section 2.2; needed for the dIoU = IoU identity (Lemma 1).
  • ad hoc to paper Mixed-operator path estimates obtained by taking max/min over exclusive-path estimates are admissible.
    Section 2.3.2 and Appendix E.3 assert this; Appendix F proves admissibility only for exclusive paths.
  • standard math Best-first search with admissible estimates and frontier pruning yields a complete search.
    Classical A*-style search theory; standard result, though not explicitly cited.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Guaranteed Optimal Compositional Explanations for Neurons." pith.science (2026). https://pith.science/paper/KHSG4KOL

@misc{pith2026251120934,
  author       = {Pith},
  title        = {Pith review of: Guaranteed Optimal Compositional Explanations for Neurons},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KHSG4KOL}},
  note         = {Machine review of arXiv:2511.20934}
}
read the original abstract

Compositional explanations are a family of methods that aim to describe the spatial alignment between neurons' receptive field activations and concepts through logical rules, typically computed via a search over all possible concept combinations. Since computing the spatial alignment over the entire state space is computationally infeasible, the literature commonly adopts assumptions related to the structure of the combinations and beam search to restrict the state space. However, beam search cannot provide any theoretical guarantees of optimality, and it remains unclear how close current explanations are to the true optimum. In this theoretical paper, we address this gap by introducing the first framework for computing guaranteed optimal compositional explanations over the entire state space spanned by the adopted assumptions. Specifically, we propose: (i) a decomposition that identifies the factors influencing the spatial alignment, (ii) a heuristic to estimate the alignment at any stage of the search, and (iii) the first algorithm that can compute optimal compositional explanations in a time comparable to exhaustive beam search. Using this framework, we demonstrate that 10-40% of explanations previously obtained with beam search are suboptimal when overlapping concepts are involved. Finally, we evaluate a beam-search variant guided by our proposed decomposition and heuristic, showing that it matches or improves runtime over prior methods while offering greater flexibility in hyperparameters and computational resources.

Figures

Figures reproduced from arXiv: 2511.20934 by the authors.

Figure 1
Figure 1. An example of a case where the optimal algorithm finds a combination of concepts [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Alignment detected in units of a ResNet18 model by both the optimal and beam-search [PITH_FULL_IMAGE:figures/full_fig_p031_2.png] view at source ↗
Figure 3
Figure 3. Alignment detected in units of a ResNet18 model by both the optimal and beam-search [PITH_FULL_IMAGE:figures/full_fig_p032_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Alignment detected in units of a ResNet18 model by both the optimal and beam-search [PITH_FULL_IMAGE:figures/full_fig_p033_4.png]
Figure 5
Figure 5. Figure 5: Alignment detected in units of an AlexNet model by both the optimal and beam-search [PITH_FULL_IMAGE:figures/full_fig_p034_5.png]
Figure 6
Figure 6. Figure 6: Alignment detected in units of an AlexNet model by both the optimal and beam-search [PITH_FULL_IMAGE:figures/full_fig_p035_6.png]
Figure 7
Figure 7. Figure 7: Alignment detected in units of an AlexNet model by both the optimal and beam-search [PITH_FULL_IMAGE:figures/full_fig_p036_7.png]
Figure 8
Figure 8. Figure 8: Alignment detected in units of a DenseNet161 model by both the optimal and beam-search [PITH_FULL_IMAGE:figures/full_fig_p037_8.png]
Figure 9
Figure 9. Figure 9: Alignment detected in units of a DenseNet161 model by both the optimal and beam-search [PITH_FULL_IMAGE:figures/full_fig_p038_9.png]
Figure 10
Figure 10. Figure 10: Alignment detected in units of a DenseNet161 model by both the optimal and beam-search [PITH_FULL_IMAGE:figures/full_fig_p039_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 4 canonical work pages

  1. [1]

    Y . H. Ahn, H. B. Kim, and S. T. Kim. Www: a unified framework for explaining what where and why of neural networks by interpretation of neuron concepts. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10968–10977, 2024

    2024

  2. [2]

    D. Bau, B. Zhou, A. Khosla, A. Oliva, and A. Torralba. Network dissection: Quantifying interpretability of deep visual representations. InComputer Vision and Pattern Recognition, 2017

    2017

  3. [3]

    Bau, J.-Y

    D. Bau, J.-Y . Zhu, H. Strobelt, A. Lapedriza, B. Zhou, and A. Torralba. Understanding the role of individual units in a deep neural network.Proceedings of the National Academy of Sciences,

  4. [4]

    Bricken, A

    T. Bricken, A. Templeton, J. Batson, B. Chen, A. Jermyn, T. Conerly, N. Turner, C. Anil, C. Deni- son, A. Askell, R. Lasenby, Y . Wu, S. Kravec, N. Schiefer, T. Maxwell, N. Joseph, Z. Hatfield- Dodds, A. Tamkin, K. Nguyen, B. McLean, J. E. Burke, T. Hume, S. Carter, T. Henighan, and C. Olah. Towards monosemanticity: Decomposing language models with dictio...

    2023

  5. [5]

    Bykov, L

    K. Bykov, L. Kopf, S. Nakajima, M. Kloft, and M. Höhne. Labeling neural representations with inverse recognition. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Neural Information Processing Systems, volume 36, pages 24804–24828. Curran Associates, Inc., 2023. URL https://proceedings.neurips.cc/paper_files/ pap...

    2023

  6. [6]

    Casper, T

    S. Casper, T. Rauker, A. Ho, and D. Hadfield-Menell. Sok: Toward transparent AI: A survey on interpreting the inner structures of deep neural networks. InFirst IEEE Conference on Secure and Trustworthy Machine Learning, 2023. URL https://openreview.net/forum? id=8C5zt-0Utdn

    2023

  7. [7]

    Cordts, M

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele. The cityscapes dataset for semantic urban scene understanding. InProc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  8. [8]

    Dreyer, E

    M. Dreyer, E. Purelku, J. Vielhaben, W. Samek, and S. Lapuschkin. Pure: Turning polysemantic neurons into pure features by identifying relevant circuits.arXiv preprint arXiv:2404.06453, 2024

    arXiv 2024

Show all 42 references
  1. [9]

    Elhage, T

    N. Elhage, T. Hume, C. Olsson, N. Schiefer, T. Henighan, S. Kravec, Z. Hatfield-Dodds, R. Lasenby, D. Drain, C. Chen, R. Grosse, S. McCandlish, J. Kaplan, D. Amodei, M. Wattenberg, and C. Olah. Toy models of superposition.Transformer Circuits Thread, 2022

    2022

  2. [10]

    Erhan, Y

    D. Erhan, Y . Bengio, A. Courville, and P. Vincent. Visualizing higher-layer features of a deep network. 2009

    2009

  3. [11]

    L. Gao, T. D. la Tour, H. Tillman, G. Goh, R. Troll, A. Radford, I. Sutskever, J. Leike, and J. Wu. Scaling and evaluating sparse autoencoders. InThe Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=tcsZt9ZNKD

    2025

  4. [12]

    L. H. Gilpin, D. Bau, B. Z. Yuan, A. Bajwa, M. Specter, and L. Kagal. Explaining explanations: An overview of interpretability of machine learning. In2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA). IEEE, oct 2018. doi: 10.1109/dsaa.2018. 00018

    2018 doi

  5. [13]

    Harth.Understanding Individual Neurons of ResNet Through Improved Compositional Formulas, pages 283–294

    R. Harth.Understanding Individual Neurons of ResNet Through Improved Compositional Formulas, pages 283–294. Springer International Publishing, 2022. ISBN 9783031092824. doi: 10.1007/978-3-031-09282-4_24. 11

    2022 doi

  6. [14]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778. IEEE, 2016

    2016

  7. [15]

    Hernandez, S

    E. Hernandez, S. Schwettmann, D. Bau, T. Bagashvili, A. Torralba, and J. Andreas. Natural lan- guage descriptions of deep features. InInternational Conference on Learning Representations,

  8. [16]

    Hesse, J

    R. Hesse, J. Fischer, S. Schaub-Meyer, and S. Roth. Disentangling polysemantic channels in convolutional neural networks. InThe First Workshop on Mechanistic Interpretability for Vision, 2025

    2025

  9. [17]

    Huang, Z

    G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger. Densely connected convolutional networks. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017

    2017

  10. [18]

    Kalibhat, S

    N. Kalibhat, S. Bhardwaj, C. B. Bruss, H. Firooz, M. Sanjabi, and S. Feizi. Identifying interpretable subspaces in image representations. InInternational Conference on Machine Learning, pages 15623–15638. PMLR, 2023

    2023

  11. [19]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convo- lutional neural networks. In F. Pereira, C. Burges, L. Bottou, and K. Weinberger, edi- tors,Advances in Neural Information Processing Systems, volume 25. Curran Associates, Inc., 2012. URL ...

    2012

  12. [20]

    B. La Rosa. Explaining deep neural networks by leveraging intrinsic methods, 2024

    2024

  13. [21]

    La Rosa, L

    B. La Rosa, L. H. Gilpin, and R. Capobianco. Towards a fuller understanding of neurons with clustered compositional explanations. InThirty-seventh Conference on Neural Information Processing Systems, 2023. URLhttps://openreview.net/forum?id=51PLYhMFWz

    2023

  14. [22]

    S. M. Makinwa, B. La Rosa, and R. Capobianco. Detection accuracy for evaluating composi- tional explanations of units. InAIxIA 2021 - Advances in Artificial Intelligence, pages 550–563. Springer International Publishing, 2022. doi: 10.1007/978-3-031-08421-8_38

    2021 doi

  15. [23]

    Massidda and D

    R. Massidda and D. Bacciu. Knowledge-driven interpretation of convolutional neural net- works. InMachine Learning and Knowledge Discovery in Databases, pages 356–371. Springer International Publishing, 2023. doi: 10.1007/978-3-031-26387-3_22

    2023 doi

  16. [24]

    Mu and J

    J. Mu and J. Andreas. Compositional explanations of neurons. InProceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20, Red Hook, NY , USA, 2020. Curran Associates Inc. ISBN 9781713829546

    2020

  17. [25]

    Nguyen, J

    A. Nguyen, J. Yosinski, and J. Clune. Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks.Visualization for Deep Learning workshop, ICML 2016, Feb. 2016

    2016

  18. [26]

    Oikarinen and T.-W

    T. Oikarinen and T.-W. Weng. CLIP-dissect: Automatic description of neuron representations in deep vision networks. InInternational Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=iPWiwWHc1V

    2023

  19. [27]

    Oikarinen and T.-W

    T. Oikarinen and T.-W. Weng. Linear explanations for individual neurons. InForty-first International Conference on Machine Learning, 2024. URL https://openreview.net/ forum?id=WIbntm28cM

    2024

  20. [28]

    C. Olah, A. Mordvintsev, and L. Schubert. Feature visualization.Distill, 2(11), nov 2017. doi: 10.23915/distill.00007

    2017 doi

  21. [29]

    O’Mahony, V

    L. O’Mahony, V . Andrearczyk, H. Müller, and M. Graziani. Disentangling neuron representa- tions with concept vectors. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3769–3774, 2023. 12

    2023

  22. [31]

    A. A. Srinivas, T. Oikarinen, D. Srivastava, W.-H. Weng, and T.-W. Weng. Sand: Enhancing open-set neuron descriptions through spatial awareness. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2993–3002. IEEE, Feb. 2025. doi: 10.1109/ wacv610...

    2025

  23. [32]

    Wang, W.-N

    A. Wang, W.-N. Lee, and X. Qi. Hint: Hierarchical neuron concept explainer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10254–10264, June 2022

    2022

  24. [33]

    Y . Wu, A. Kirillov, F. Massa, W.-Y . Lo, and R. Girshick. Detectron2.https://github.com/ facebookresearch/detectron2, 2019

    2019

  25. [34]

    B. Zhou, A. Lapedriza, A. Khosla, A. Oliva, and A. Torralba. Places: A 10 million image database for scene recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017

    2017

  26. [35]

    B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba. Scene parsing through ade20k dataset. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, July 2017. doi: 10.1109/cvpr.2017.544. 13 Table 3: Visualization of identified quantities f...

    2017 doi

  27. [38]

    All results were computed on a workstation equipped with an NVIDIA GTX 3090 GPU, without parallelization, in order to avoid timing overhead

    ( accessible at: https://www.cityscapes-dataset.com/ under MIT License) for the low complexity settings, Ade20kFull [35] (accessible via the Detectron2 [33] framwork) for the moderate settings, and Broden [2] (accessible at https://github.com/CSAILVision/NetDissect under MIT l...

  28. [39]

    Compute the exact quantities (Section 2.2) for every concept in the dataset (line 5)

  29. [40]

    For each concept, compute dIoU max and dIoU min for all possible paths starting from it, using the heuristic aggregated computation (line 6). 15

  30. [41]

    The frontier is sorted by the estimateddIoU max

    Initialize the frontier with all paths (line 7) and keep track of the greatest dIoU min (line 8). The frontier is sorted by the estimateddIoU max

  31. [42]

    Reduce the frontier by removing nodes whose estimated dIoU max is lower than the global maximum of the minimum estimates (line 10)

  32. [43]

    final path

    Iteratively pop nodes from the frontier, starting from those with the highest estimated dIoU (lines 11-12). (a) If the estimate is aggregated, refine it by computing the sample-based estimate and reinsert the node (lines 13-20). Otherwise, proceed to the next step. (b) Apply l...

    2000

  33. [2020]

    doi: 10.1073/pnas.1907375117

    ISSN 0027-8424. doi: 10.1073/pnas.1907375117. URL https://www.pnas.org/ content/early/2020/08/31/1907375117

    2020 doi

  34. [2022]

    URLhttps://openreview.net/forum?id=NudBMY-tzDr

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.