Pith. sign in

REVIEW 3 major objections 6 minor 44 references

InfoCons: Identifying Interpretable Critical Concepts in Point Clouds via Information Theory

T0 review · 3 major / 6 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 InfoCons finds the points that drive a point-cloud prediction.

desk verdict A practical and broad empirical study of VIB-style attribution for point clouds, but the central information-theoretic derivation does not hold up, so treat the theory as a heuristic and the experiments as the contribution. read the letter →

arxiv 2505.19820 v1 pith:JGJTQR2N submitted 2025-05-26 cs.LG

Feifei Li , Mi Zhang , Zhaoxiang Wang , Min Yang This is my paper
classification cs.LG
keywords pointcloudinterpretabilitycriticalsubsetinformationbottleneckmutualattributionmethodsconcept-levelexplanationpoint-dropattack3Ddeeplearning
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

InfoCons tries to answer a concrete question about a point cloud classifier: which points, taken together, actually caused this prediction? The paper's central claim is that a critical subset that is both faithful and conceptually coherent can be obtained by maximizing a variational lower bound on the mutual information between the subset and the model's output, while penalizing the information shared with the input. The authors instantiate this as a learned soft mask over the encoder's intermediate features, with a stop-gradient Gaussian substitution that prevents unimportant points from being penalized for carrying neighbor information. The result, they argue, is a score map whose highest-scoring points form recognizable object parts, and whose removal degrades accuracy more than removing baseline-selected points.

What carries the argument

The explanatory object is the attention bottleneck $f(\hat m \mid z(x); \theta)$: a channel-wise attention block that reads the frozen encoder's intermediate feature $z\in\mathbb{R}^{D\times N'}$, projects it through linear layers with ELU, applies a scaled softmax attention, and outputs a mask $\hat m\in(0,1)^{D\times N'}$ via an MLP with sigmoid. The mask is optimized under Eq. 6, whose first term is the cross-entropy lower bound on $I(C,y)$ and whose second term, the KL divergence between $\hat z$ and a Gaussian prior $q(\hat z)=\mathcal{N}(\mu_z,\sigma_z^2)$ parameterized by the feature's own mean and variance, is the variational proxy for the information-bottleneck penalty $I(x,C)$. The stop-gradient noise substitution $\hat z=\hat m\odot z+\operatorname{sg}(1-\hat m)\odot\epsilon$ is the piece that lets the mask point to the anchor point's unique information instead of the entangled neighbor context.

What would settle it

For a small model and a single point cloud, randomly sample many subsets of the same size as the InfoCons subset and record how much prediction changes when each is dropped; if random subsets routinely change predictions far more than the InfoCons subset, the claim that Eq. 6 identifies the most faithful minimal subset is falsified.

Watch

Extended reading notes

Core claim

InfoCons proposes that an interpretable critical subset of a point cloud is one that maximizes $I(C, y)$ — the mutual information between the selected points $C$ and the model's prediction $y$ — while minimizing $I(x, C)$, the information the selection still carries about the full input $x$. The paper rewrites this information-bottleneck objective for a pretrained point cloud model as $$\max_\$\theta$ \mathbb{E}_{x\sim p(x)}\left[\mathbb{E}_{y\sim p(y)}\log q(y|\hat z) - \$\beta$ D_{KL}(\hat z \| q(\hat z))\right], \quad \hat z = \hat m \odot z(x) + \operatorname{sg}(1-\hat m)\odot \epsilon,$$ where $\hat m = f(\hat m \mid z(x); \theta)$ is a learned soft mask produced by an attention bottleneck, $z(x)$ is the intermediate feature of the frozen encoder, and $\epsilon_i \sim \mathcal{N}(\mu_z, \sigma_z^2)$ is Gaussian noise matched to the feature statistics. The stop-gradient term is the paper's mechanism for disentangling each point's own information from the information it shares with its neighbors: when a point is deemed unimportant, its feature is replaced by neighbor-typical noise rather than by zero, so the model is only penalized for removing the point's unique contribution. Once trained, averaging the soft mask over the reduced dimension gives a per-point critical score $s(x)$, and the critical subset is the set of highest-scoring points.

Load-bearing premise

The load-bearing premise is that the KL divergence computed on intermediate features, $D_{KL}(\hat z \| q(\hat z))$, is a valid proxy for the input-level bottleneck $I(x,C)$, so that minimizing it really enforces that the selected subset is minimal.

Editorial extensions

If this is right

  • If Eq. 6 does what the paper claims, explaining a point cloud model reduces to a one-pass forward computation, yielding 200-point critical subsets that drop DGCNN test accuracy to 63.70% after 20 iterations, second only to a 1,000-query black-box baseline.
  • The same score map is reusable as a saliency signal: replacing SageMix's gradient saliency with InfoCons raises DGCNN mixup test accuracy from 92.79% to 93.19%, and rescaling SI-Adv's gradient with InfoCons increases attack success rate while lowering Chamfer and Hausdorff distances.
  • Because the mask is trained on the frozen model's features rather than on logits alone, the method transfers across non-hierarchical, hierarchical, and self-attention architectures, with evaluation on eight models, and extends to object detection by substituting a detection-score objective for classification cross-entropy.
  • K-means clustering of the InfoCons score map produces a critical-subset hierarchy, letting a user decompose a point cloud into levels such as 'contour' versus 'seat' for a sofa misclassified as a bench.

Reading between the lines

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

  • A natural test the paper does not run: measure how well InfoCons's top-scoring points agree with human part annotations such as PartNet labels; high agreement would show conceptual coherence is an emergent property of the objective rather than a visualization artifact.
  • The stop-gradient Gaussian substitution is a general recipe for attribution under feature entanglement; porting it to 2D image or video attributions, where neighboring pixels share features, could remove the same i.i.d. failure that motivated Eq. 6.
  • The paper's manual choice of intermediate layer $l$ and $\beta$ invites a data-driven criterion: select layer and $\beta$ that maximize the MCD-LCD accuracy gap on a validation set, rather than score-map variance.
  • If the KL proxy ever fails to track $I(x,C)$, InfoCons would still be a useful saliency heuristic; the paper's own observation that features are entangled, with nearly all points highlighted for hierarchical models, is the evidence to watch.
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 / 6 minor

Summary. The paper introduces InfoCons, a post-hoc explanation method for point cloud classifiers. InfoCons trains a small attention-based bottleneck module on a frozen model's intermediate features; the module outputs a soft mask that is applied to the features, optionally replacing non-critical feature entries with Gaussian noise. The training objective (Eq. 6) combines the cross-entropy of the masked features with the model's prediction and an information-loss term D_KL(z_hat||q(z_hat)), presented as a variational approximation of the Information Bottleneck objective I(C,y)-beta*I(x,C). The resulting scores are used to extract critical subsets, which are evaluated with point-drop attacks on eight architectures and three datasets, and are plugged into SageMix data augmentation and SI-Adv adversarial attack. The paper also provides qualitative comparisons with CP, CP++, PCSAM, and LIME3D.

Significance. If the IB derivation were valid, InfoCons would be a useful contribution: it is a single-forward-pass explainer that scales across a wide range of point cloud architectures, and the downstream applications (mixup and adversarial attack) show measurable gains. The empirical study is extensive (8 architectures, 3 datasets, 4 baselines) and the code is released. However, the central theoretical claim is currently unsupported: Eq. (6) does not follow from the variational bound in Eq. (9), and the information-loss term as written is not the KL divergence claimed. The method may still work as a heuristic feature-replacement regularizer, and the empirical evidence is promising, but the paper needs to either supply a correct bound/derivation or explicitly reframe the contribution as heuristic. The qualitative concept-coherence evidence is compelling but only anecdotal.

major comments (3)
  1. [§3.2, Eq. (6)] The information-loss term D_KL(z_hat||q(z_hat)) in Eq. (6) is not a well-defined KL for the distribution induced by the construction. With z_hat = m_hat * z + sg(1 - m_hat) * epsilon and epsilon_i ~ N(mu_z, sigma_z^2), the conditional distribution of each dimension is p(z_hat_i | x) = N(m_hat_i z_i + (1 - m_hat_i) mu_i, (1 - m_hat_i)^2 sigma_i^2). The exact KL against q = N(mu_z, sigma_z^2) contains -log(1 - m_hat_i), which diverges as m_hat_i -> 1. The paper does not state which finite approximation is used in the implementation; without this, the objective in Eq. (6) is not the variational penalty claimed, and the reported training numbers cannot be checked against the formula.
  2. [§A.2, Eqs. (15)-(17)] The step E_x[D_KL(p(z_hat|x)||q(z_hat))] = E_x[D_KL(z_hat||q(z_hat))] is a notational substitution and does not establish an inequality. Moreover, the variational upper bound in Eq. (9) applies to I(x,z_hat), the mutual information between the input and the feature-replaced representation, not to I(x,C) for the selected input subset C = m * x. Equation (17) similarly asserts I(C,x) <= E_x[D_KL(m_hat||q(m_hat))] without proof. Consequently the central claim that Eq. (6) optimizes the IB objective in Eq. (2) for critical concepts is not supported by the derivation; the method is at present a heuristic feature-replacement regularizer for which the information-theoretic grounding remains to be supplied.
  3. [§4, evaluation protocol] Conceptual coherence, one of the two central desiderata in the paper's title and introduction, is supported only qualitatively (Fig. 6, Fig. 11, Fig. 15). The only quantitative evaluation is the point-drop attack, which is a faithfulness proxy and is itself acknowledged in §4.4 to have limitations (removing points in specific spatial locations can push the point cloud off the data manifold). No quantitative measure of concept coherence is provided, so the claim that InfoCons produces critical subsets that are both faithful and conceptually coherent is not fully evidenced.
minor comments (6)
  1. [§3.2, Eq. (6)] q(z_hat) is called a 'Gaussian prior' but is parameterized with the same feature statistics (mu_z, sigma_z^2) used in the noise insertion, making it a data-dependent prior; the term 'learnable unbiased prior' in the abstract is not defined.
  2. [§3.2, Eq. (7)] The softmax axis of Att(q_z,z,v_z) is not specified; q_z^T z has shape (N,N) and the result is not obviously a per-point attention weight. Please clarify the normalization and the output shape.
  3. [§4.2] The abbreviation 'PCSM' is used after introducing 'PCSAM' (and the baseline is called PCSAM elsewhere); please make the notation consistent.
  4. [§C.3, Eq. (18)] The expectation E_k is not defined; use an explicit expectation over the Gumbel noise variables and specify the noise distribution clearly.
  5. [§4.3, Table 2] The abbreviations CD and HD (Chamfer distance and Hausdorff distance, presumably) should be defined at first use.
  6. [Table 4] The header 'Sparsity/% P I(z=0)/(D' x N')' is hard to parse; define P and the sparsity statistic explicitly.

Circularity Check

1 steps flagged · score 2.0 of 10

No construction-level circularity: point-drop, baseline, and application evaluations are external and self-contained, and there are no self-citations. The theoretical IB-grounding of Eq. 6 is definitional, since Appendix A.2 (Eq. 16) equates the required bound on I(x,C) with the implemented loss by notation.

  1. self definitional [Appendix A.2, Eq. 16; Section 3.2, Definition 3.2 (Eq. 6)]
    "Ex[DKL(p(ẑ|x)||q(ẑ))] = Ex[DKL(ẑ||q(ẑ))] considering ẑ=f(ẑ|x;θ). Moreover, since z(x) is determined by the trained encoder F, we consider the upper bound based on m̂ in Eq. 5 as follows: I(C,x) ≤ Ex[DKL(m̂||q(m̂))]."

    The paper claims (Def. 3.2, Eq. 6) that Eq. 6 is the IB objective for Eq. 2 (max I(C,y) − βI(x,C)). In App. A.2 the variational upper bound on I(x,C) is written as Ex[DKL(p(ẑ|x)||q(ẑ))] and then equated to the per-sample loss Ex[DKL(ẑ||q(ẑ))] by the phrase 'considering ẑ=f(ẑ|x;θ)'. This equality is a notational substitution, not a derived bound: with the induced p(ẑ|x)=N(m̂⊙z,(1−m̂)²σ²) and q=N(µ_z,σ²_z), the exact KL contains −log(1−m̂) and diverges as m̂→1, so no finite KL of the stated distributions equals the implemented loss. The claim that Eq. 6 minimizes I(x,C) thus holds by definition of the loss, not by the IB derivation; the empirical drop-attack evaluation is external and unaffected.

full rationale

Walking the claimed derivation chain (Eq. 2 → Eq. 5 → Eq. 6 and Appendix A.2): the paper contains no self-citations, so the self-citation patterns do not apply. The explainer θ is trained with cross-entropy on masked features plus a KL regularizer; its score maps are evaluated on a held-out test split by point-drop attack (MCD/LCD) and against four external baselines (CP, CP++, PCSAM, LIME3D), with additional downstream metrics (mixup test OA, ASR, CD/HD) that do not appear in the training loss. No fitted constant is renamed as a prediction: the score map is a genuine function of learned parameters, and the faithfulness results are not forced by the optimization. One step in the theoretical narrative is definitional: Appendix A.2 (Eq. 16) equates the required variational upper bound on I(x,C) with the implemented loss DKL(ẑ||q(ẑ)) by the phrase 'considering ẑ=f(ẑ|x;θ)', so the claim that Eq. 6 is the IB objective for Eq. 2 is true by notation rather than by derivation; this is the same gap the skeptic identifies (the exact KL of the induced Gaussian would diverge as m̂→1). This weakens the information-theoretic grounding but does not collapse any empirical prediction. Minor caveats: Section 4.4 selects β 'together with the test set accuracy' on the same test split used for the reported drop-attack numbers, a mild evaluation-integrity concern rather than construction-level circularity; and conceptual coherence is supported by qualitative evidence plus the KL regularization, not by the unestablished I(x,C) bound. On balance, the paper's empirical contribution is self-contained against external benchmarks, and the circularity score is accordingly low.

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

The core IB bounds (Eqs. 8-9) are standard. The method adds free hyperparameters (beta, D_r, layer l, Gumbel tau) that are tuned or manually set per model, and relies on the assumptions that feature-space masking equals input-space critical subsets and that point features are i.i.d. for KL estimation. No new physical or conceptual entities are introduced beyond the learned mask.

free parameters (4)
  • beta = 10^k, varied per model (e.g., 10^-3 to 10^3)
    Weights the information loss in Eq. 5/6; chosen by grid search on validation (Fig. 9) and reported as critical to performance.
  • reduction dimension D_r = 64 (default)
    Dimension of the attention bottleneck (Eq. 7); controls the prior distribution; set by parameter study.
  • intermediate layer l = manually selected per model (Table 4)
    Determines which encoder layer's features z are masked; the text says 'we manually determine the intermediate layer for eight models' (Appendix C.4).
  • Gumbel softmax temperature tau = 0.7
    Controls mask sharpness in Eq. 18; affects the score map smoothness.
assumptions (4)
  • standard math Variational bounds on I(z,y) and I(z,x) (Eqs. 8-9) hold.
    Used to derive the VIB objective in Eq. 10, following Alemi et al. (2016).
  • domain assumption A soft mask on point features is equivalent to a critical subset of input points.
    Stated after Eq. 4: 'm_hat optimized for z is also applicable for x'. This assumes feature-space masking faithfully represents input-space selection.
  • domain assumption Point features are i.i.d. for Monte Carlo KL estimation.
    Appendix A.2: 'we assume that each point feature is i.i.d.' This contradicts the paper's own motivation that features are entangled.
  • domain assumption The model's output Y is the correct target signal for training the explainer.
    The cross-entropy loss in Eq. 5/6 uses a label y; it is implicit (not fully specified) whether this is the ground-truth or the model's prediction, which matters for explaining misclassifications.

how reviews work

0 comments
Cite this review

Pith. "Pith review of InfoCons: Identifying Interpretable Critical Concepts in Point Clouds via Information Theory." pith.science (2026). https://pith.science/paper/JGJTQR2N

@misc{pith2026250519820,
  author       = {Pith},
  title        = {Pith review of: InfoCons: Identifying Interpretable Critical Concepts in Point Clouds via Information Theory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JGJTQR2N}},
  note         = {Machine review of arXiv:2505.19820}
}
read the original abstract

Interpretability of point cloud (PC) models becomes imperative given their deployment in safety-critical scenarios such as autonomous vehicles. We focus on attributing PC model outputs to interpretable critical concepts, defined as meaningful subsets of the input point cloud. To enable human-understandable diagnostics of model failures, an ideal critical subset should be *faithful* (preserving points that causally influence predictions) and *conceptually coherent* (forming semantically meaningful structures that align with human perception). We propose InfoCons, an explanation framework that applies information-theoretic principles to decompose the point cloud into 3D concepts, enabling the examination of their causal effect on model predictions with learnable priors. We evaluate InfoCons on synthetic datasets for classification, comparing it qualitatively and quantitatively with four baselines. We further demonstrate its scalability and flexibility on two real-world datasets and in two applications that utilize critical scores of PC.

Figures

Figures reproduced from arXiv: 2505.19820 by the authors.

Figure 1
Figure 1. Attributing PC model outputs to a group of interpretable critical concepts using InfoCons. The derived concepts, which conform to a specific semantically meaningful structure (i.e., con￾ceptual cohesion), can reflect their influence on the model outputs faithfully. The critical score map can also be integrated into multi￾ple applications. Better viewed in color. safety-critical scenarios like autonomous vehicles (Ge… view at source ↗
Figure 2
Figure 2. (a) Theoretical illustration for extracting critical concepts. (b) Overview of our explanation framework to obtain the InfoCons. PC model for shape classification is denoted as G ◦ F. SymFunc stands for symmetric functions (e.g., maxpooling). Attention Bottleneck θ is trained with end-to-end objectives. Once trained, θ uses only the intermediate feature to provide explanations. corners, observed across various point… view at source ↗
Figure 4
Figure 4. InfoCons-based critical subsets (200 pts) for four PC models (covering three distinct structure types) are compared with meanpool-based CP++ and gradient-based PCSAM. PCSAM tends to extract similar and spatially aggregated subsets for all models, while InfoCons identifies more interpretable critical subsets that are faithful to model behavior (color distinction not required). tive concepts for a given PC model can d… view at source ↗
Figures from the paper (15 more)
Figure 5
Figure 5. Figure 5: The data flow of the bottleneck network f(·|z(x); θ). The input z is the intermediate feature from F, and the output mˆ holds the same dimension as z. the hierarchical models and self-attention based models. We introduce a non-linear attention block to learn the unbias…
Figure 6
Figure 6. Figure 6: Qualitative comparisons of InfoCons with four baselines (CP, CP++, PCSAM, LIME3D) include both score maps (best viewed in color) and critical subsets. The PCs are selected from ModelNet40, where DGCNN gives incorrect predictions. out the least critical points (LCD). We…
Figure 7
Figure 7. Figure 7: (B,C). Since PCSM is a gradient-based method that utilize the ground truth label to calculate the loss, our Info￾Cons may not outperform PCSAM when dropping only a small critical subset. However, our method InfoCons (1pass) achieves better performance when remove a lar…
Figure 8
Figure 8. Figure 8: InfoCons-based mixup (left) and adversarial attack (right). InfoCons-based score map can also explain the effec￾tiveness of adversarial examples. Better viewed in color. based saliency score map of SageMix (Lee et al., 2022) with InfoCons, then following SageMix’s dyna…
Figure 9
Figure 9. Figure 9: Impacts of β, Dr and intermediate layer l for DGCNN@l on ModelNet40. 𝛽 = 10! 𝛽∗ = 10# 𝛽 = 10$ 𝛽 = 0 * 𝛽 = 0 means ℒ!" only Original Detections of PointPillars A Missing Car! Why? 60 steps (Eq.6) Explanation for Missing Objects ∗Example: KITTI Test Data (id=007517) Targ…
Figure 10
Figure 10. Figure 10: Impacts of β for DGCNN@5 on ScanObjectNN (left) and the explanation for missing objects of object detector PointPil￾lars@1 (right, better viewed in color). The implementation details can be found in Appendix C.2. mentation and adversarial attacks. The results are deri…
Figure 11
Figure 11. Figure 11: Decompose point cloud redundancy by visualizing the Critical Points Hierarchy [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Additional samples for Critical Points Hierarchy. clustering (K = 4). We observe that the hierarchy of In￾foCons suggests the most critical concept is the contour of a ‘sofa’ (as the model incorrectly predicts), and the least critical concept is the seat (a common att…
Figure 13
Figure 13. Figure 13: Dynamic critical subset (200 points) with one-pass and multi-iteration score maps, compared with PCSAM. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 14
Figure 14. Figure 14: Failure cases for PCT@4 with varying βk ∈ {101 /k, 102 /k, 103 /k, 104 /k}, where k = N × D and α with Dr = D/α. D.2. Additional Qualitative Results • In [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: InfoCons on the PC labeled as ‘monitor’ for 8 PC models (from left to right: (i)the 1024 points score map, (ii)the dynamic critical points with 10 iterations and 10 points dropped each iteration, 100 points in total, (iii) dynamic critical points with 20 iteration, 20…
Figure 16
Figure 16. Figure 16: InfoCons compared with two baselines on DGCNN (Wang et al., 2019), and we demonstrate 200 points critical subset. From left to right, (i) InfoCons, (ii) Critical Points++ (Levi & Gilboa, 2023), (iii)-(v) PC Saliency Map (Zheng et al., 2019) with a dynamic process of p…
Figure 17
Figure 17. Figure 17: Effectiveness evaluation of InfoCons on eight PC models by point-drop attack, where 10 to 500 points are dropped. We report the changes of accuracy under two settings: (i) dropping out the most critical points (MCD) and (ii) dropping out the least critical points (LCD…
Figure 18
Figure 18. Figure 18: Additional comparisons of InfoCons and PC Saliency Map for [PITH_FULL_IMAGE:figures/full_fig_p018_18.png]
Figure 19
Figure 19. Figure 19: Quantitative comparisons between our proposed objectives (Eq. 5 and Eq. 6) and Critical Points++ (Levi & Gilboa, 2023) for eight PC models. We observe an abnormal trend for PointNet++, where CP++ assigns critical points as unimportant, resulting in a rapidly dropping …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 34 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]

    A., Fischer, I., Dillon, J

    Alemi, A. A., Fischer, I., Dillon, J. V., and Murphy, K. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016

    arXiv 2016

  3. [3]

    R., Jiang, H., Brilakis, I., Fischer, M., and Savarese, S

    Armeni, I., Sener, O., Zamir, A. R., Jiang, H., Brilakis, I., Fischer, M., and Savarese, S. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 1534--1543, 2016

    work page 2016

  4. [4]

    Explaining a black-box by using a deep variational information bottleneck approach

    Bang, S., Xie, P., Lee, H., Wu, W., and Xing, E. Explaining a black-box by using a deep variational information bottleneck approach. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pp.\ 11396--11404, 2021

    work page 2021

  5. [5]

    and Bethge, M

    Brendel, W. and Bethge, M. Approximating cnns with bag-of-local-features models works surprisingly well on imagenet. In International Conference on Learning Representations, 2019

    work page 2019

  6. [6]

    X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al

    Chang, A. X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015

    arXiv 2015

  7. [7]

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

    Chen, J., Song, L., Wainwright, M., and Jordan, M. Learning to explain: An information-theoretic perspective on model interpretation. In International conference on machine learning, pp.\ 883--892. PMLR, 2018

    work page 2018

  8. [8]

    Unifying fourteen post-hoc attribution methods with taylor interactions

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

    work page 2024

Show all 44 references
  1. [9]

    Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds

    Feng, T., Quan, R., Wang, X., Wang, W., and Yang, Y. Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 1761--1769, 2024

    2024

  2. [10]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Geiger, A., Lenz, P., and Urtasun, R. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pp.\ 3354--3361. IEEE, 2012

    2012

  3. [11]

    R., and Hu, S.-M

    Guo, M.-H., Cai, J.-X., Liu, Z.-N., Mu, T.-J., Martin, R. R., and Hu, S.-M. Pct: Point cloud transformer. Computational Visual Media, 7: 0 187--199, 2021

    2021

  4. [12]

    Towards better visualizing the decision basis of networks via unfold and conquer attribution guidance

    Hong, J.-H., Nam, W.-J., Jeon, K.-S., and Lee, S.-W. Towards better visualizing the decision basis of networks via unfold and conquer attribution guidance. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp.\ 7884--7892, 2023

    2023

  5. [13]

    Shape-invariant 3d adversarial point clouds

    Huang, Q., Dong, X., Chen, D., Zhou, H., Zhang, W., and Yu, N. Shape-invariant 3d adversarial point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 15335--15344, 2022

    2022

  6. [14]

    Minimal adversarial examples for deep learning on 3d point clouds

    Kim, J., Hua, B.-S., Nguyen, T., and Yeung, S.-K. Minimal adversarial examples for deep learning on 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 7797--7806, 2021

    2021

  7. [15]

    W., Nguyen, T., Tang, Y

    Koh, P. W., Nguyen, T., Tang, Y. S., Mussmann, S., Pierson, E., Kim, B., and Liang, P. Concept bottleneck models. In International conference on machine learning, pp.\ 5338--5348. PMLR, 2020

    2020

  8. [16]

    H., Vora, S., Caesar, H., Zhou, L., Yang, J., and Beijbom, O

    Lang, A. H., Vora, S., Caesar, H., Zhou, L., Yang, J., and Beijbom, O. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 12697--12705, 2019

    2019

  9. [17]

    Lee, S., Jeon, M., Kim, I., Xiong, Y., and Kim, H. J. Sagemix: Saliency-guided mixup for point clouds. Advances in Neural Information Processing Systems, 35: 0 23580--23592, 2022

    2022

  10. [18]

    Levi, M. Y. and Gilboa, G. Critical points++: An agile point cloud importance measure for robust classification, adversarial defense and explainable ai. arXiv preprint arXiv:2308.05525, 2023

    2023 arXiv

  11. [19]

    B., and Saligrama, V

    Lin, F., Yue, Y., Zhang, Z., Hou, S., Yamada, K., Kolachalama, V. B., and Saligrama, V. Infocd: A contrastive chamfer distance loss for point cloud completion. In Thirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  12. [20]

    Liu, H., Cai, M., and Lee, Y. J. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pp.\ 657--675. Springer, 2022

    2022

  13. [21]

    Ffam: Feature factorization activation map for explanation of 3d detectors

    Liu, S., Li, B., Fang, Z., Cui, M., and Huang, K. Ffam: Feature factorization activation map for explanation of 3d detectors. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024 a

    2024

  14. [22]

    Point-voxel cnn for efficient 3d deep learning

    Liu, Z., Tang, H., Lin, Y., and Han, S. Point-voxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems, 32, 2019

    2019

  15. [23]

    Timex++: Learning time-series explanations with information bottleneck

    Liu, Z., Wang, T., Shi, J., Zheng, X., Chen, Z., Song, L., Dong, W., Obeysekera, J., Shirani, F., and Luo, D. Timex++: Learning time-series explanations with information bottleneck. In Forty-first International Conference on Machine Learning, 2024 b

    2024

  16. [24]

    Rethinking network design and local geometry in point cloud: A simple residual mlp framework

    Ma, X., Qin, C., You, H., Ran, H., and Fu, Y. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022

    2022 arXiv

  17. [25]

    Curvenet: Curvature-based multitask learning deep networks for 3d object recognition

    Muzahid, A., Wan, W., Sohel, F., Wu, L., and Hou, L. Curvenet: Curvature-based multitask learning deep networks for 3d object recognition. IEEE/CAA Journal of Automatica Sinica, 8 0 (6): 0 1177--1187, 2020

    2020

  18. [26]

    R., Su, H., Mo, K., and Guibas, L

    Qi, C. R., Su, H., Mo, K., and Guibas, L. J. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 652--660, 2017 a

    2017

  19. [27]

    R., Yi, L., Su, H., and Guibas, L

    Qi, C. R., Yi, L., Su, H., and Guibas, L. J. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017 b

    2017

  20. [28]

    A practical review of mechanistic interpretability for transformer-based language models

    Rai, D., Zhou, Y., Feng, S., Saparov, A., and Yao, Z. A practical review of mechanistic interpretability for transformer-based language models. arXiv preprint arXiv:2407.02646, 2024

    2024

  21. [29]

    why should i trust you?

    Ribeiro, M. T., Singh, S., and Guestrin, C. " why should i trust you?" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pp.\ 1135--1144, 2016

    2016

  22. [30]

    M., and Bischof, H

    Schinagl, D., Krispel, G., Possegger, H., Roth, P. M., and Bischof, H. Occam's laser: Occlusion-based attribution maps for 3d object detectors on lidar data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 1141--1150, 2022

    2022

  23. [31]

    Restricting the flow: Information bottlenecks for attribution

    Schulz, K., Sixt, L., Tombari, F., and Landgraf, T. Restricting the flow: Information bottlenecks for attribution. arXiv preprint arXiv:2001.00396, 2020

    2001 arXiv

  24. [32]

    R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., and Batra, D

    Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., and Batra, D. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pp.\ 618--626, 2017

    2017

  25. [33]

    Visualizing global explanations of point cloud dnns

    Tan, H. Visualizing global explanations of point cloud dnns. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp.\ 4741--4750, 2023

    2023

  26. [34]

    and Kotthaus, H

    Tan, H. and Kotthaus, H. Surrogate model-based explainability methods for point cloud nns. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp.\ 2239--2248, 2022

    2022

  27. [35]

    C., and Bialek, W

    Tishby, N., Pereira, F. C., and Bialek, W. The information bottleneck method. arXiv preprint physics/0004057, 2000

    2000 arXiv

  28. [36]

    A., Pham, Q.-H., Hua, B.-S., Nguyen, T., and Yeung, S.-K

    Uy, M. A., Pham, Q.-H., Hua, B.-S., Nguyen, T., and Yeung, S.-K. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 1588--1597, 2019

    2019

  29. [37]

    E., Bronstein, M

    Wang, Y., Sun, Y., Liu, Z., Sarma, S. E., Bronstein, M. M., and Solomon, J. M. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38 0 (5): 0 1--12, 2019

    2019

  30. [38]

    Point transformer v3: Simpler faster stronger

    Wu, X., Jiang, L., Wang, P.-S., Liu, Z., Liu, X., Qiao, Y., Ouyang, W., He, T., and Zhao, H. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 4840--4851, 2024

    2024

  31. [39]

    3d shapenets: A deep representation for volumetric shapes

    Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., and Xiao, J. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 1912--1920, 2015

    1912

  32. [40]

    R., and Li, B

    Xiang, C., Qi, C. R., and Li, B. Generating 3d adversarial point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 9136--9144, 2019

    2019

  33. [41]

    R., Guibas, L., and Litany, O

    Xie, S., Gu, J., Guo, D., Qi, C. R., Guibas, L., and Litany, O. Pointcontrast: Unsupervised pre-training for 3d point cloud understanding. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part III 16, pp.\ 574--591. Springer, 2020

    2020

  34. [42]

    Learning geometry-disentangled representation for complementary understanding of 3d object point cloud

    Xu, M., Zhang, J., Zhou, Z., Xu, M., Qi, X., and Qiao, Y. Learning geometry-disentangled representation for complementary understanding of 3d object point cloud. In AAAI, 2021

    2021

  35. [43]

    Pointcloud saliency maps

    Zheng, T., Chen, C., Yuan, J., Li, B., and Ren, K. Pointcloud saliency maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 1598--1606, 2019

    2019

  36. [44]

    Explaining generalization power of a dnn using interactive concepts

    Zhou, H., Zhang, H., Deng, H., Liu, D., Shen, W., Chan, S.-H., and Zhang, Q. Explaining generalization power of a dnn using interactive concepts. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 17105--17113, 2024

    2024

Pith tools

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