acceptodds
Under review as a conference paper at ICLR 2027

CLAG-SNN: Compact Local, Abstract Global for Efficient Spiking 3D Point Cloud Recognition

Abstract

Spiking neural networks (SNNs) offer prominent advantages in efficient 3D point-cloud recognition. However, the discrete nature of spiking representations and the irregular structure of point clouds make it difficult to simultaneously preserve local geometry and form high-level semantic abstractions. Inspired by the hierarchical processing mechanism of biological vision, we propose a **Compact Local, Abstract Global (CLAG)** principle, which preserves nonlinear dependencies before and after compression at the local scale to maintain structure, and suppresses nonlinear dependencies between inputs and spiking features at the global scale to promote abstraction. We instantiate this principle as a hierarchical framework termed CLAG-SNN. First, at the local scale, we propose **Hierarchical Structure-aware Token Compaction (HSTC)**, which preserves local geometric dependencies through structure-aware token aggregation and compaction, achieving local compactness. Furthermore, at the global scale, we propose **Nonlinearity-Aware Abstraction Bottleneck (NAAB)**, which suppresses redundant input–feature nonlinear coupling and abstracts compact tokens into discriminative global semantics. Experiments on multiple 3D point-cloud recognition benchmarks show that CLAG-SNN achieves consistent improvements over the baseline: while improving accuracy, it increases inference speed by up to 24.4% and reduces theoretical energy consumption by up to one-third; with only 1% training data, it still achieves a 3.32% performance gain.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks, Neuromorphic Computing, Event-driven Computing

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_uAMZOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes the "Compact Local, Abstract Global" (CLAG) representation principle for 3D point cloud representation using SNNs. Originally, this idea is inspired by The Hilbert–Schmidt Independence Criterion as a unified metric of nonlinear dependency. The author proposed two complementary methods: HSTC, which performs token compression based on SVD spectral analysis and an adaptive token budget, together with an HSIC consistency constraint; and NAAB, which performs HSIC dependency suppression together with a diversity mechanism anchored by entropy and kernel density. Experiments cover classification, segmentation, and detection tasks on five datasets: ModelNet10/40, ScanObjectNN, SemanticKITTI, and KITTI, as long as report low-data and ablation results. The paper also provides duality/complementary slackness analysis, generalization bounds, and sample complexity analysis.

Critical strengths
  1. Using HSIC as a unified measure of nonlinear dependency and pointing out that dependency should be regulated in opposite directions at the local and global scales is more reasonable than "uniform compression" or "uniform preservation", which are usuallt applied by many prior works. Proofs show that HSTC combines spectral analysis, adaptive budgeting, saliency, and batch-level structural consistency. While NAAB simultaneously considers dependency suppression and representation diversity, avoiding collapse caused by merely minimizing HSIC. (Importance: 9/10)
  2. The supplementary analysis regarding the duality and generalization bounds of the proposed CLAG contributes positively to the study of representation learning in SNNs. (Importance: 8/10)
  3. Experiments show that this approach is quite powerful for 3D tasks, beating up the trending methods like E-3DSNN and SpikingPointMamba. (Importance: 6/10)
Critical weaknesses
  1. Regarding the relationship between local enhancement of nonlinear dependence and high-level reduction of nonlinear dependence and the biological visual system, the author only uses the hierarchical structure of the primary visual cortex as an analogy, without delving into the functional organization mechanism of the relevant division of labor in the visual nervous system. (Importance: 9/10)
  2. Neither the manuscript nor the supp. material provides experimental visualization results for the actual model, making it difficult to intuitively gauge the model's performance. (Importance: 8/10)
Questions for authors

Clarification requests:

  1. Please provide more biological evidence to support the core idea of the paper: Is bio-CNS appear similar behavior that preserves non linear dependencies before and after compression at the local scale, and suppresses it between inputs and spiking features at the global scale?

Suggestions for this and future work:

  1. The authors did not provide a comparison of model parameter counts in the experiments section, just accuracy, FPS and theoretical energy, which is inappropriate for fair comparison.
  2. The layout of Figure 1 is overly cluttered, and the curves on the right-hand side are not explained in the figure legend. Furthermore, Figure 1 is not cited anywhere in the main text.
  3. The ablation study only reports results, but the efficiency differences between “HSTC only” and “NAAB only” are not explained in the main text. It is recommended to add analysis of these differences.
  4. Typos that needs checking: inconsistently uses “3D point-cloud” and “3D point cloud” in the title and main text;
  5. Recommended to add a unified notation table in the appendix. The paper introduces a large number of symbols, such as BB,NbN_b,TT and DD. Although some symbols are defined when they first appear, they are scattered across the main text and appendix, resulting in poor readability.
3 comments
AuthorsOpenReview ↗

Response to Reviewer uAMZ: Strengthened Biological Grounding, Added Visualizations, and Fair Parameter Comparison

We sincerely thank the reviewer for providing the most constructive feedback among the three reviews. Your insightful comments have significantly helped us improve the quality of this manuscript.

W1 & C1: We have expanded the theoretical foundation by incorporating functional theories such as Predictive Coding and Information Bottleneck, rather than relying solely on anatomical analogies. We have refined the Introduction to explicitly state that CLAG is a computational principle. Additionally, we have refined the theoretical derivations.

W2 : We have added a t-SNE visualization in Appendix J (Fig. 3) to intuitively demonstrate the evolution from entangled input representations to highly separated latent semantic clusters, effectively showcasing the performance of our method.

S1 :We have added a parameter comparison. CLAG-SNN has 1,875,536 parameters, while the E-3DSNN baseline has 1,874,216. The increase of exactly 1,320 parameters is solely due to the auxiliary classification head. HSTC is parameter-free, and NAAB's only trainable addition is this auxiliary head. The modules change where computation is spent, not the parameter count.

S3 :We have added a detailed analysis of the ablation study in the main text. HSTC physically reduces the token count to achieve a substantial FPS gain. NAAB does not alter the token count but reduces power consumption and improves accuracy by suppressing redundancy. The combination achieves both spatial and temporal efficiency.

S2, S4, S5 : We have completely redrawn Figure 1, properly cited it in the text with detailed legends, added a comprehensive notation table in Appendix I (Table 13), and thoroughly proofread the manuscript to standardize terminology (e.g., "3D point cloud").

We believe these rigorous revisions have substantially improved the theoretical soundness, experimental transparency, and practical value of this work. We sincerely thank the reviewer for the guidance that elevated the quality of this paper.

Reviewer_uAMZOpenReview ↗

Response

I appreciate the authors‘ response. When I said that I asked for more visualization results, I actually mean the visualization of the Segmentation tasks, since the visual aid speaks for itself. I'll raise my score for most of my concerns are addressed in the new version manuscript.

AuthorsOpenReview ↗

Response to Reviewer uAMZ

We sincerely thank the reviewer for the clarification. We apologize for the misunderstanding regarding the visualization results. We fully agree that visual comparisons speak for themselves for the semantic segmentation task on SemanticKITTI. We will include these visualizations in our revised manuscript. We greatly appreciate your constructive suggestions.

Reviewer_36gZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes the CLAG principle of locally preserving dependencies and globally compressing dependencies. The local module HSTC determines the number of reserved tokens based on the SVD spectrum of the feature matrix, selects tokens through significance scores, and maximizes the HSIC of the descriptors before and after compression; The global module NAAB minimizes the HSIC between input descriptors and latent variables, while incorporating entropy maximization and KDE regularization. The author applies it to point cloud classification, segmentation, and detection, and claims to achieve a joint improvement in accuracy, speed, and theoretical energy consumption. The paper proposes the CLAG principle of locally preserving dependencies and globally compressing dependencies. The local module HSTC determines the number of reserved tokens based on the SVD spectrum of the feature matrix, selects tokens through significance scores, and maximizes the HSIC of the descriptors before and after compression; The global module NAAB minimizes the HSIC between input descriptors and latent variables, while incorporating entropy maximization and KDE regularization. The author applies it to point cloud classification, segmentation, and detection, and claims to achieve a joint improvement in accuracy, speed, and theoretical energy consumption.

Critical strengths
  1. The problem is valuable, and it is intuitively reasonable to use different constraints for local and global situations.
  2. Conducted basic component and robustness analysis
Critical weaknesses
  1. Proof of the generalization theorem is not valid. Assumption 3 in Appendix B directly assumes that semantic distortion is controlled by the input feature HSIC, and compression distortion is controlled by the reciprocal of the local HSIC. These are the core relationships that need to be explained or verified in the paper, and we cannot simply claim that the theory reveals the validity of the proposed principles based on their assumptions. The main problem is that the proof of the generalization theorem is not valid. Assumption 3 in Appendix B directly assumes that semantic distortion is controlled by the input feature HSIC, and compression distortion is controlled by the reciprocal of the local HSIC. These are the core relationships that need to be explained or verified in the paper, and we cannot simply claim that the theory reveals the validity of the proposed principles based on their assumptions.

  2. The connection between HSIC constraints, local geometric preservation, and global semantic abstraction is insufficient. Equation (7) calculates cross sample HSIC for the average token descriptor of each sample, without directly constraining adjacency relationships, relative positions, or local shapes. The descriptor dependency is strong and insufficient to ensure the preservation of geometric structures. The input descriptor of NAAB only represents the mean of coordinates and attributes, and cannot fully represent the structural information of the original point cloud.

Questions for authors
  1. How to revise the estimation bounds of cross model risk difference in the appendix and derive the sample complexity of Theorem 2?

  2. What are the actual sample size, bandwidth calculation, and gradient processing of HSIC on SemanticKITTI?

  3. After taking into account SVD, Top-K, and integer pulse amplitudes, does the complete inference cost and accuracy speed trade-off still maintain an advantage?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 36gZ: Re-establishing Generalization Validity, Clarifying Semantic Boundaries, and Computational Analysis

We deeply appreciate the reviewer's rigorous and thorough review. We highly value the critical issues raised by the reviewer regarding theoretical proofs, mechanism consistency, and experimental interpretability. We have thoroughly revised the manuscript by: completely rewriting the theoretical proofs, adding extensive control experiments, and carefully calibrating our claims to ensure transparency.

W1 & Q1: We sincerely appreciate the reviewer for pointing out this fundamental flaw. We fully agree and have restated Theorem 1 as a conditional generalization bound. There are three major modifications:

(i) Explicit definitions of all objects. Let H={g∘fθ:θ∈Θ}\mathcal{H}=\{g\circ f_\theta:\theta\in\Theta\}, where Θ\Theta is compact (bounded parameter norm); let h⋆∈arg⁡min⁡h∈HR(h)h^\star\in\arg\min_{h\in\mathcal{H}}R(h) be the best-in-class hypothesis, and the ideal representation be the latent mapping induced by h⋆h^\star. The loss is truncated as ℓM=min⁡{ℓ,M}\ell_M=\min\{\ell,M\}, ensuring 0≤ℓM≤M0\le\ell_M\le M and maintaining LℓL_\ell-Lipschitzness (since truncation is 1-Lipschitz); LgL_g is the spectral norm of the final linear map in the prediction head. Consequently, the boundedness (Assumption 1, with Bz=1B_z=1 after the L2 normalization described in Q2) and Lipschitzness hold by construction.

(ii) Decomposition explicitly isolating the between-hypothesis term.

We decompose the error as:

R(h)−R^S(h)=[R(h)−R(h⋆)]+[R(h⋆)−R^S(h⋆)]+[R^S(h⋆)−R^S(h)]R(h) - \widehat{R}_S(h) = [R(h) - R(h^\star)] + [R(h^\star) - \widehat{R}_S(h^\star)] + [\widehat{R}_S(h^\star) - \widehat{R}_S(h)]

where the first term is the representation error, the second is the estimation error of h-star, and the third is the uniform control term. The third bracket is bounded by:

2sup⁡h∈H∣R(h)−R^S(h)∣≤2Rn(H)+2log⁡(2/δ)/2n2 \sup_{h \in \mathcal{H}} |R(h) - \hat{R}_S(h)| \le 2 \mathcal{R}_n(\mathcal{H}) + 2\sqrt{\log(2/\delta)/2n}

so h and h* are covered by the same event — this is exactly the control you asked for. The revised bound is:

R(h)−R^S(h)≤LlLgΔZ(h)+2Rn(H)+3log⁡(2/δ)/2nR(h) - \widehat{R}_S(h) \le L_l L_g \Delta_Z(h) + 2 R_n(H) + 3\sqrt{\log(2/\delta)/2n}

which holds simultaneously for all hypotheses in the class, with the Rademacher complexity bounded by:

Rn(H)≤CHlog⁡n/n\mathcal{R}_n(\mathcal{H}) \le C_H \sqrt{\log n / n}

for bounded spectral norms. The complexity is that of the hypothesis class; we no longer derive a data-manifold-dependent rate from the input dimension alone.

(iii) Graded labeling of assumptions. Assumption 3 is rewritten as a modeling assumption ("not a derived fact") within the hypothesis environment itself, with both sides computed during training; Section 4.3 reports the five terms for the trained model, and Table 5 reports the matched-retention controls. Assumption 4 (entropy-controlled covering) is replaced by a standard covering bound implied by bounded latents: log⁡N(ϵ)≤dzlog⁡(1+2Bz/ϵ)\log\mathcal{N}(\epsilon)\le d_z\log(1+2B_z/\epsilon) with Bz=1B_z=1. Assumption 5 is merged into the same modeling decomposition. Theorem 2 is rewritten in a bias-variance form: ΩCLAG\Omega_{\mathrm{CLAG}} is a non-vanishing bias term independent of nn; under the checkable condition ΩCLAG≤ϵ/(2LℓLg)\Omega_{\mathrm{CLAG}}\le\epsilon/(2L_\ell L_g), the gap is at most ϵ\epsilon as long as n≥32C2ϵ2[dHlog⁡32C2ϵ2+log⁡2δ]n\ge\frac{32C^2}{\epsilon^2}[d_{\mathcal{H}}\log\frac{32C^2}{\epsilon^2}+\log\frac{2}{\delta}] (where dHd_{\mathcal{H}} is the covering dimension of H\mathcal{H}). The step multiplying ΓCLAG\Gamma_{\mathrm{CLAG}} into the covering number has been removed, and terms like "guarantee / proves / reveals" have been deleted throughout the text.

We have refined Sections 3.3 and 3.4 of the main text, as well as Appendix B.

Q2: SemanticKITTI is trained with a batch size of 2, which leads to a rank deficiency in the centered Gram matrix (rank ≤B−1=1\le B-1 = 1), causing the HSIC estimation to degrade. Therefore, we employ a 64-sample FIFO descriptor buffer: descriptors are accumulated across iterations, and the HSIC loss is computed on this buffer, resulting in an effective sample size of 64 instead of 2. No gradient accumulation is involved. We have added this implementation detail to the main text.

Q3: The SVD complexity is O(min⁡(Nbdl2,Nb2dl))O(\min(N_b d_l^2, N_b^2 d_l)) per sample, applied to the compacted token set. It only runs during training: inference uses a calibrated fixed retention rate (r≈0.73r\approx0.73) and skips SVD entirely. This makes the method deployable: SVD is merely a training-time tool, and the inference path is SVD-free and hardware-friendly.

AuthorsOpenReview ↗

Response to Reviewer 36gZ: Re-establishing Generalization Validity, Clarifying Semantic Boundaries, and Computational Analysis

W2 : We clarify that "local" in this paper specifically refers to the intermediate latent feature space, rather than the raw XYZ coordinates. To address your concerns regarding "whether HSIC truly captures geometric structure" and "whether the constraint mechanism is effective", we have added the following detailed experiments and statistical profiles: a. We added matched-retention control experiments in Table 5 to address whether random/native downsampling is sufficient and whether the HSIC direction matters. We introduced 4 control groups (random selection, fixed-prefix, magnitude-only, and reversed-HSIC) and evaluated them across multiple random seeds, proving the necessity of direction regulation in the CLAG principle. b. We added geometric fidelity metrics in Table 5 to address whether mean descriptors lose local geometry. We directly measured the Chamfer distance and Coverage between retained and full token sets. HSTC does not outperform random sampling in physical spatial coverage. This result precisely proves that HSTC is not a uniform spatial downsampler; its constraint operates on the relational distribution of descriptors (semantic level), rather than original geometric isometry. c. We added descriptor statistics in Table 6 to address whether the input descriptor degenerates due to centered XYZ and whether the latent variable collapses or suffers from scale explosion. We analyzed the channel variances and between/within-class variance ratios. Conclusion: The input descriptor xinx^{in} contains 62.5% degenerate channels (encoding low-order nuisance statistics), rather than a constantly zero XYZ mean; while the latent zz has 0% degenerate channels and a between/within-class variance ratio of 1.24, proving that under NAAB constraints, the features neither collapse nor scale-explode, and maintain strong class discriminability.

DatasetMethodAccuracy (%) ↑
ModelNet10E-3DSNN (AAAI)93.24 ± 0.46
ModelNet10CLAG-SNN (Ours)94.08 ± 0.22
ModelNet40E-3DSNN (AAAI)90.61 ± 0.26
ModelNet40CLAG-SNN (Ours)91.38 ± 0.09
Strategyr=0.5 Chamfer ↓r=0.5 Cov@2 ↑r=0.5 Acc (%)r=0.7 Chamfer ↓r=0.7 Cov@2 ↑r=0.7 Acc (%)
CLAG-SNN (HSTC)0.7710.90590.66 ± 0.140.3780.98291.38 ± 0.09
Random selection0.6150.97690.48 ± 0.240.3390.99591.11 ± 0.17
Fixed first-k0.6150.97688.86 ± 0.120.3390.99589.60 ± 0.02
Magnitude-only0.7800.89488.36 ± 0.050.3770.97990.86 ± 0.08
Reversed HSIC directions0.9740.85490.24 ± 0.080.4590.94990.73 ± 0.33
DescriptorDimMean stdMax stdDegenerateBetween/within
x^in160.0790.23962.5%1.20
z320.6490.8690.0%1.24

We believe these rigorous revisions have substantially improved the theoretical soundness, experimental transparency, and practical value of our work. We sincerely thank the reviewer for the guidance that helped elevate the quality of this work.

AuthorsOpenReview ↗

Response to Reviewer 36gZ: Re-establishing Generalization Validity, Clarifying Semantic Boundaries, and Computational Analysis

We have revised Appendix B, and the modifications are as follows:

We first prove Proposition 2, whose argument is self-contained, and then state the assumptions used in Theorem 1 and prove Theorem 1 and Theorem 2.

Proof of Proposition 2.

For (i), rearrange Eq. (11) as: $ L_{total} = L_{task} - \lambda_{HSTC}(L_{HSTC} - \varepsilon_{loc}) + \lambda_{NAAB}(L_{NAAB} - \varepsilon_{glb}) + C $, where $ C = \lambda_{NAAB}\varepsilon_{glb} - \lambda_{HSTC}\varepsilon_{loc} $ is constant once the prices are fixed. For feasible $ \theta $, the two slacks satisfy $ L_{HSTC} - \varepsilon_{loc} \ge 0 $ and $ L_{NAAB} - \varepsilon_{glb} \le 0 $, so each price term is non-positive and $ L_{total}(\theta) \le L_{task}(\theta) + C $. Minimizing over $ \theta $ preserves the inequality against the constrained optimum, which yields the stated lower bound; tightening it by choosing the multipliers is exactly dual maximization, which is why we call them prices.

For (ii), at a saddle point a positive price requires its constraint to be tight, since otherwise the Lagrangian could be raised by lowering that price; each price-slack product in Eq. (12) therefore vanishes.

Assumption 1 (Compact manifold and bounded representations). The input distribution is supported on a compact $ d_{\mathrm{eff}} $-dimensional manifold $ \mathcal{M} $, and the latent descriptors satisfy $ |\mathbf{z}_b|_2 \le B_z $ for all samples; after the L2 normalisation of Section 3.3, $ B_z = 1 $ by construction.

Assumption 2 (Lipschitz prediction). The loss $ \ell(\cdot,y) $ is $ L_\ell $-Lipschitz in its first argument and is clipped to $ [0,M] $; the prediction head $ g $ is $ L_g $-Lipschitz, with $ L_g $ equal to the spectral norm of its final linear map. Both constants are checkable.

Assumption 3 (Modelling decomposition of representation error). There exist constants $ c_1, c_2, c_3, c_4, c_5 > 0 $ such that the discrepancy between the learned latent representation and the best-in-class representation satisfies: $ \Delta_Z \le c_1\sqrt{L_{dep}} + c_2/\sqrt{H(Z)+\varepsilon_0} + c_3 L_{KDE} + c_4/\sqrt{L_{HSTC}+\varepsilon_t} + c_5(1-r_{keep}) $. This is a modelling assumption, not a fact derived from the architecture: all five quantities are computable during training, and Section 4.3 reports them together with the measured gap. In the main configuration the entropy and KDE anchors are ablated, so the corresponding $ c_2, c_3 $ terms are inactive.

Assumption 3 plays the role formerly assigned to a separate perturbation lemma: it is assumed and measured, not derived. The estimation term in Theorem 1 is proved separately by uniform convergence.

Proof of Theorem 1.

Fix $ h \in H $ and add and subtract $ R(h^{\ast}) $ and $ \hat{R}_{S}(h^{\ast}) $:

=[R(h)-R(h^\star)]+[R(h^\star)-\widehat R_S(h^\star)] +[\widehat R_S(h^\star)-\widehat R_S(h)].$$ The first bracket is at most \$ L_{\ell} L_{g} \Delta_{Z}(h) \$ by Assumption 2 and the definition of \$ \Delta_{Z} \$. Since \$ \ell_{M} \in [0,M] \$, Hoeffding's inequality gives: $$|R(h^{\ast}) - \hat{R}_{S}(h^{\ast})| \le M\sqrt{\log(2/\delta)/2n}$$ with probability at least \$ 1-\delta/2 \$; the constant \$ M \$ is absorbed below. For the third bracket: $$|\widehat R_S(h^\star)-\widehat R_S(h)|\le2\sup_{h\in\mathcal H}|R(h)-\widehat R_S(h)|$$ Symmetrization followed by McDiarmid's inequality gives, with probability at least \$ 1-\delta/2 \$: $$\sup_{h\in\mathcal H}|R(h)-\widehat R_S(h)|\le2\mathfrak R_n(\mathcal H)+\sqrt{\log(2/\delta)/(2n)}$$ A union bound over the two events and absorption of constants yields Eq. (16). **Theorem 2 (Sample complexity: bias--variance form).** Fix \$ \epsilon > 0 \$. Suppose the trained parameters satisfy the checkable bias condition: \$ \Omega_{CLAG}(h) := \sqrt{L_{dep}} + 1/\sqrt{H(Z)+\varepsilon_0} + L_{KDE} + 1/\sqrt{L_{HSTC}+\varepsilon_t} + (1-r_{keep}) \le \epsilon/(2L_\ell L_g) \$. Then \$ R(h)-\hat{R}_S(h) \le \epsilon \$ with probability at least \$ 1-\delta \$ whenever: \$ n \ge (32C^2/\epsilon^2) [ d_H\log(32C^2/\epsilon^2) + \log(2/\delta) ] \$, where \$ d_H \$ is the covering dimension of \$ \mathcal{H} \$. **Proof of Theorem 2.** By Theorem 1 and Eq. (17), \$ R(h)-\hat{R}_S(h) \le \epsilon/2 + 2C\sqrt{d_H\log n/n} + 3\sqrt{\log(2/\delta)/(2n)} \$. Requiring the two estimation terms to be at most \$ \epsilon/2 \$ and solving for \$ n \$ gives Eq. (18).
Reviewer_oy7VOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose CLAG-SNN, which incorporates HSTC and NAAB into a hierarchical spiking point cloud network. HSTC determines how many tokens to retain based on the singular value spectrum of the feature matrix, selects tokens according to activation strength and their relationship to the global descriptor, and maximizes HSIC between the descriptors before and after compression. At the final layer, NAAB reduces HSIC between the input descriptor and latent features while incorporating entropy and kernel density terms, aiming to obtain more abstract representations without collapse. The paper also presents a generalization bound and reports classification, detection, and segmentation results on ModelNet10/40, ScanObjectNN, KITTI, and SemanticKITTI.

Critical strengths
  • On ModelNet40, a significant percentage-point accuracy gain is accompanied; gains are also reported on SemanticKITTI and KITTI detection. For tasks that require processing sparse point clouds under resource constraints, this combination is more meaningful than improving accuracy solely by increasing model capacity.
  • The paper evaluates classification, semantic segmentation, and object detection, and provides ablation evidence suggesting that the modules may be complementary. The appendix also discloses the datasets, training configurations, and limitations of the theoretical energy consumption estimates.
Critical weaknesses
  • Writing. Setting aside the method itself, the paper introduces numerous module acronyms that are difficult to remember, making it challenging to follow the manuscript. I encourage the authors to use more memorable or meaningful acronyms.
  • The generalization argument is not adequately established: Assumption 3 in Appendix B directly assumes that HSIC correlates with semantic/structural error in the desired directions, which is precisely the key property that needs justification. More specifically, the proof of Theorem 1 treats R(h∗)−R^S(h)R(h^*)-\hat R_S(h) as an estimation error that can be directly controlled by a standard concentration bound, even though the two terms concern different hypotheses; the additional term R^S(h∗)−R^S(h)\hat R_S(h^*)-\hat R_S(h) generally also needs to be addressed. Theorem 2 then moves from a bound containing a CΩCLAGC\Omega_{\mathrm{CLAG}} term that does not vanish with sample size to a pure 1/n1/\sqrt{n}-type bound, without deriving how that term is eliminated or controlled. The current proofs are therefore insufficient to turn the training objective into a generalization guarantee.
  • The regularization does not yet align with the claimed geometric and abstraction mechanisms: The local HSIC term operates on each sample’s mean descriptor rather than local adjacency relations, while the global input descriptor is the mean of coordinates and attributes. Different geometries can share the same mean, and centered XYZ-only inputs may even have identical zero means. These quantities therefore may not capture the claimed local shape or irrelevant details. Section 3.3 also claims that the kernel density term limits excessive dispersion, but its summation includes each sample itself, so this term is at most log⁡B\log B and cannot, on its own, offset Gaussian entropy that can grow with scaling. The existing module ablations are insufficient to distinguish effective empirical regularization from the specific mechanisms described in the paper.
Questions for authors
  • Gaps in the proofs. Please clarify the definitions of the ideal representation and h∗h^*, the hypothesis class, and the boundedness conditions on the loss; complete the control of the empirical risk difference between different hypotheses in Theorem 1; and explain how Theorem 2 handles CΩCLAGC\Omega_{\mathrm{CLAG}}. Can Assumptions 3–5 be derived from verifiable network conditions? If not, should the theory be framed as a conditional explanation rather than a proven generalization mechanism?
  • Scale constraints. Does the implementation impose an explicit norm constraint or normalization on ZZ? If not, how does LKDE\mathcal{L}_{\mathrm{KDE}}, which includes self-similarity terms, prevent entropy from increasing through feature scaling? How is the boundedness in Assumption 1 of the appendix ensured during training?
  • Informative content of the descriptors. Are the inputs centered for each dataset, what attributes does xinx^{\mathrm{in}} contain, and how much within-batch variance does it actually exhibit? Under the same retention rate and training settings, can controls using fixed/random pruning, a mean-preservation loss, and reversed local/global HSIC directions support the proposed mechanism? These controls directly determine whether the CLAG principle holds, rather than constituting general future work.
  • Reliability and scope of the gains. Please report variability across multiple random seeds, actual token retention ratios, complete hyperparameters, and end-to-end runtime and computational costs including SVD and sorting. Does the energy consumption formula include the added SVD operations, and how does it account for integer-valued spike amplitudes?
6 comments
AuthorsOpenReview ↗

Response to Reviewer oy7V: Resolving Theoretical Flaws, Clarifying Geometric Mechanism, and Adding Rigorous Controls

We sincerely thank the reviewer for the detailed, insightful, and highly constructive feedback. We deeply value the critical issues raised regarding theoretical proofs, mechanism consistency, and experimental interpretability. In this revision, we have completely rewritten the theoretical proofs, added extensive control experiments, and carefully calibrated our claims.

W1: We have added a comprehensive notation table in Appendix I (Table 13) and standardized the terminology throughout the manuscript.

W2 & Q1: We sincerely appreciate the reviewer for pointing out this fundamental flaw. We fully agree and have restated Theorem 1 as a conditional generalization bound. There are three major modifications:

(i) Explicit definitions of all objects. Let H={g∘fθ:θ∈Θ}\mathcal{H}=\{g\circ f_\theta:\theta\in\Theta\}, where Θ\Theta is compact (bounded parameter norm); let h⋆∈arg⁡min⁡h∈HR(h)h^\star\in\arg\min_{h\in\mathcal{H}}R(h) be the best-in-class hypothesis, and the ideal representation be the latent mapping induced by h⋆h^\star. The loss is truncated as ℓM=min⁡{ℓ,M}\ell_M=\min\{\ell,M\}, ensuring 0≤ℓM≤M0\le\ell_M\le M and maintaining LℓL_\ell-Lipschitzness (since truncation is 1-Lipschitz); LgL_g is the spectral norm of the final linear map in the prediction head. Consequently, the boundedness (Assumption 1, with Bz=1B_z=1 after the L2 normalization described in Q2) and Lipschitzness hold by construction.

(ii) Decomposition explicitly isolating the between-hypothesis term.

We decompose the error as:

R(h)−R^S(h)=[R(h)−R(h⋆)]+[R(h⋆)−R^S(h⋆)]+[R^S(h⋆)−R^S(h)]R(h) - \widehat{R}_S(h) = [R(h) - R(h^\star)] + [R(h^\star) - \widehat{R}_S(h^\star)] + [\widehat{R}_S(h^\star) - \widehat{R}_S(h)]

where the first term is the representation error, the second is the estimation error of h-star, and the third is the uniform control term. The third bracket is bounded by:

2sup⁡h∈H∣R(h)−R^S(h)∣≤2Rn(H)+2log⁡(2/δ)/2n2 \sup_{h \in \mathcal{H}} |R(h) - \hat{R}_S(h)| \le 2 \mathcal{R}_n(\mathcal{H}) + 2\sqrt{\log(2/\delta)/2n}

so h and h* are covered by the same event — this is exactly the control you asked for. The revised bound is:

R(h)−R^S(h)≤LlLgΔZ(h)+2Rn(H)+3log⁡(2/δ)/2nR(h) - \widehat{R}_S(h) \le L_l L_g \Delta_Z(h) + 2 R_n(H) + 3\sqrt{\log(2/\delta)/2n}

which holds simultaneously for all hypotheses in the class, with the Rademacher complexity bounded by:

Rn(H)≤CHlog⁡n/n\mathcal{R}_n(\mathcal{H}) \le C_H \sqrt{\log n / n}

for bounded spectral norms. The complexity is that of the hypothesis class; we no longer derive a data-manifold-dependent rate from the input dimension alone.

(iii) Graded labeling of assumptions. Assumption 3 is rewritten as a modeling assumption ("not a derived fact") within the hypothesis environment itself, with both sides computed during training; Section 4.3 reports the five terms for the trained model, and Table 5 reports the matched-retention controls. Assumption 4 (entropy-controlled covering) is replaced by a standard covering bound implied by bounded latents: log⁡N(ϵ)≤dzlog⁡(1+2Bz/ϵ)\log\mathcal{N}(\epsilon)\le d_z\log(1+2B_z/\epsilon) with Bz=1B_z=1. Assumption 5 is merged into the same modeling decomposition. Theorem 2 is rewritten in a bias-variance form: ΩCLAG\Omega_{\mathrm{CLAG}} is a non-vanishing bias term independent of nn; under the checkable condition ΩCLAG≤ϵ/(2LℓLg)\Omega_{\mathrm{CLAG}}\le\epsilon/(2L_\ell L_g), the gap is at most ϵ\epsilon as long as n≥32C2ϵ2[dHlog⁡32C2ϵ2+log⁡2δ]n\ge\frac{32C^2}{\epsilon^2}[d_{\mathcal{H}}\log\frac{32C^2}{\epsilon^2}+\log\frac{2}{\delta}] (where dHd_{\mathcal{H}} is the covering dimension of H\mathcal{H}). The step multiplying ΓCLAG\Gamma_{\mathrm{CLAG}} into the covering number has been removed, and terms like "guarantee / proves / reveals" have been deleted throughout the text.

We have refined Sections 3.3 and 3.4 of the main text, as well as Appendix B.

Q2. Regarding scale constraints: In our implementation, all latent-side terms are based on the L2-normalized latent descriptor z~b=zb/∥zb∥2\widetilde{z}_b=z_b/\|z_b\|_2, hence ∥z~b∥2=1\|\widetilde{z}_b\|_2=1. In every training iteration, Assumption 1 is naturally satisfied with a bounded constant Bz=1B_z=1. The normalization eliminates scale variation issues: on the unit sphere, the maximum pairwise distance is 2, and the maximum per-channel variance is 1; therefore, the Gaussian entropy has an upper bound H≤dz2log⁡2πeH \le \frac{d_z}{2}\log 2\pi e. The entropy cannot continuously grow with feature scale expansion. The self-similar KDE term is also bounded (saturating at log⁡B\log B when descriptors are far apart).

Q4. We added multi-seed experiments on ModelNet10 and ModelNet40 in Table 4 to verify the reliability of our method. The SVD complexity is O(min⁡(Nbdl2,Nb2dl))O(\min(N_b d_l^2, N_b^2 d_l)) per sample, applied to the compacted token set. It only runs during training: inference uses a calibrated fixed retention rate (r≈0.73r\approx0.73) and skips SVD entirely. This makes the method deployable: SVD is merely a training-time tool, and the inference path is SVD-free and hardware-friendly.

AuthorsOpenReview ↗

Response to Reviewer oy7V: Resolving Theoretical Flaws, Clarifying Geometric Mechanism, and Adding Rigorous Controls

W3 & Q3: We clarify that "local" in this paper specifically refers to the intermediate latent feature space, rather than the raw XYZ coordinates. To address your concerns regarding "whether HSIC truly captures geometric structure" and "whether the constraint mechanism is effective", we have added the following detailed experiments and statistical profiles: a. We added matched-retention control experiments in Table 5 to address whether random/native downsampling is sufficient and whether the HSIC direction matters. We introduced 4 control groups (random selection, fixed-prefix, magnitude-only, and reversed-HSIC) and evaluated them across multiple random seeds, proving the necessity of direction regulation in the CLAG principle. b. We added geometric fidelity metrics in Table 5 to address whether mean descriptors lose local geometry. We directly measured the Chamfer distance and Coverage between retained and full token sets. HSTC does not outperform random sampling in physical spatial coverage. This result precisely proves that HSTC is not a uniform spatial downsampler; its constraint operates on the relational distribution of descriptors (semantic level), rather than original geometric isometry. c. We added descriptor statistics in Table 6 to address whether the input descriptor degenerates due to centered XYZ and whether the latent variable collapses or suffers from scale explosion. We analyzed the channel variances and between/within-class variance ratios. Conclusion: The input descriptor xinx^{in} contains 62.5% degenerate channels (encoding low-order nuisance statistics), rather than a constantly zero XYZ mean; while the latent zz has 0% degenerate channels and a between/within-class variance ratio of 1.24, proving that under NAAB constraints, the features neither collapse nor scale-explode, and maintain strong class discriminability.

DatasetMethodAccuracy (%) ↑
ModelNet10E-3DSNN (AAAI)93.24 ± 0.46
ModelNet10CLAG-SNN (Ours)94.08 ± 0.22
ModelNet40E-3DSNN (AAAI)90.61 ± 0.26
ModelNet40CLAG-SNN (Ours)91.38 ± 0.09
Strategyr=0.5 Chamfer ↓r=0.5 Cov@2 ↑r=0.5 Acc (%)r=0.7 Chamfer ↓r=0.7 Cov@2 ↑r=0.7 Acc (%)
CLAG-SNN (HSTC)0.7710.90590.66 ± 0.140.3780.98291.38 ± 0.09
Random selection0.6150.97690.48 ± 0.240.3390.99591.11 ± 0.17
Fixed first-k0.6150.97688.86 ± 0.120.3390.99589.60 ± 0.02
Magnitude-only0.7800.89488.36 ± 0.050.3770.97990.86 ± 0.08
Reversed HSIC directions0.9740.85490.24 ± 0.080.4590.94990.73 ± 0.33
DescriptorDimMean stdMax stdDegenerateBetween/within
x^in160.0790.23962.5%1.20
z320.6490.8690.0%1.24

We believe the rigorous revisions above have substantially improved the theoretical soundness, experimental transparency, and practical value of this work. We sincerely thank the reviewer for the guidance that helped elevate the quality of this paper.

AuthorsOpenReview ↗

Response to Reviewer oy7V: Resolving Theoretical Flaws, Clarifying Geometric Mechanism, and Adding Rigorous Controls

We have revised Appendix B, and the modifications are as follows:

We first prove Proposition 2, whose argument is self-contained, and then state the assumptions used in Theorem 1 and prove Theorem 1 and Theorem 2.

Proof of Proposition 2.

For (i), rearrange Eq. (11) as: $ L_{total} = L_{task} - \lambda_{HSTC}(L_{HSTC} - \varepsilon_{loc}) + \lambda_{NAAB}(L_{NAAB} - \varepsilon_{glb}) + C $, where $ C = \lambda_{NAAB}\varepsilon_{glb} - \lambda_{HSTC}\varepsilon_{loc} $ is constant once the prices are fixed. For feasible $ \theta $, the two slacks satisfy $ L_{HSTC} - \varepsilon_{loc} \ge 0 $ and $ L_{NAAB} - \varepsilon_{glb} \le 0 $, so each price term is non-positive and $ L_{total}(\theta) \le L_{task}(\theta) + C $. Minimizing over $ \theta $ preserves the inequality against the constrained optimum, which yields the stated lower bound; tightening it by choosing the multipliers is exactly dual maximization, which is why we call them prices.

For (ii), at a saddle point a positive price requires its constraint to be tight, since otherwise the Lagrangian could be raised by lowering that price; each price-slack product in Eq. (12) therefore vanishes.

Assumption 1 (Compact manifold and bounded representations). The input distribution is supported on a compact $ d_{\mathrm{eff}} $-dimensional manifold $ \mathcal{M} $, and the latent descriptors satisfy $ |\mathbf{z}_b|_2 \le B_z $ for all samples; after the L2 normalisation of Section 3.3, $ B_z = 1 $ by construction.

Assumption 2 (Lipschitz prediction). The loss $ \ell(\cdot,y) $ is $ L_\ell $-Lipschitz in its first argument and is clipped to $ [0,M] $; the prediction head $ g $ is $ L_g $-Lipschitz, with $ L_g $ equal to the spectral norm of its final linear map. Both constants are checkable.

Assumption 3 (Modelling decomposition of representation error). There exist constants $ c_1, c_2, c_3, c_4, c_5 > 0 $ such that the discrepancy between the learned latent representation and the best-in-class representation satisfies: $ \Delta_Z \le c_1\sqrt{L_{dep}} + c_2/\sqrt{H(Z)+\varepsilon_0} + c_3 L_{KDE} + c_4/\sqrt{L_{HSTC}+\varepsilon_t} + c_5(1-r_{keep}) $. This is a modelling assumption, not a fact derived from the architecture: all five quantities are computable during training, and Section 4.3 reports them together with the measured gap. In the main configuration the entropy and KDE anchors are ablated, so the corresponding $ c_2, c_3 $ terms are inactive.

Assumption 3 plays the role formerly assigned to a separate perturbation lemma: it is assumed and measured, not derived. The estimation term in Theorem 1 is proved separately by uniform convergence.

Proof of Theorem 1.

Fix $ h \in H $ and add and subtract $ R(h^{\ast}) $ and $ \hat{R}_{S}(h^{\ast}) $:

=[R(h)-R(h^\star)]+[R(h^\star)-\widehat R_S(h^\star)] +[\widehat R_S(h^\star)-\widehat R_S(h)].$$ The first bracket is at most \$ L_{\ell} L_{g} \Delta_{Z}(h) \$ by Assumption 2 and the definition of \$ \Delta_{Z} \$. Since \$ \ell_{M} \in [0,M] \$, Hoeffding's inequality gives: $$|R(h^{\ast}) - \hat{R}_{S}(h^{\ast})| \le M\sqrt{\log(2/\delta)/2n}$$ with probability at least \$ 1-\delta/2 \$; the constant \$ M \$ is absorbed below. For the third bracket: $$|\widehat R_S(h^\star)-\widehat R_S(h)|\le2\sup_{h\in\mathcal H}|R(h)-\widehat R_S(h)|$$ Symmetrization followed by McDiarmid's inequality gives, with probability at least \$ 1-\delta/2 \$: $$\sup_{h\in\mathcal H}|R(h)-\widehat R_S(h)|\le2\mathfrak R_n(\mathcal H)+\sqrt{\log(2/\delta)/(2n)}$$ A union bound over the two events and absorption of constants yields Eq. (16). **Theorem 2 (Sample complexity: bias--variance form).** Fix \$ \epsilon > 0 \$. Suppose the trained parameters satisfy the checkable bias condition: \$ \Omega_{CLAG}(h) := \sqrt{L_{dep}} + 1/\sqrt{H(Z)+\varepsilon_0} + L_{KDE} + 1/\sqrt{L_{HSTC}+\varepsilon_t} + (1-r_{keep}) \le \epsilon/(2L_\ell L_g) \$. Then \$ R(h)-\hat{R}_S(h) \le \epsilon \$ with probability at least \$ 1-\delta \$ whenever: \$ n \ge (32C^2/\epsilon^2) [ d_H\log(32C^2/\epsilon^2) + \log(2/\delta) ] \$, where \$ d_H \$ is the covering dimension of \$ \mathcal{H} \$. **Proof of Theorem 2.** By Theorem 1 and Eq. (17), \$ R(h)-\hat{R}_S(h) \le \epsilon/2 + 2C\sqrt{d_H\log n/n} + 3\sqrt{\log(2/\delta)/(2n)} \$. Requiring the two estimation terms to be at most \$ \epsilon/2 \$ and solving for \$ n \$ gives Eq. (18).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.