acceptodds
Under review as a conference paper at ICLR 2027

COSP: Consumer Oriented Second-order Pruning for Compact Implicit Neural Representations

Abstract

Implicit neural representations (INRs) encode a signal as a coordinate network. We study post-training structured compaction, which requires identifying hidden structures that can be removed with little functional damage. This is difficult because deleting a neuron changes adjacent affine tensors, whereas conventional group scores need not match the underlying functional perturbation and globally comparable second-order scores are costly. We propose COSP (Consumer Oriented Second-order Pruning) for structured INR compaction. COSP derives a module-wise second-order back-propagation rule and its diagonal propagation variant, which evaluate candidate groups without materializing a global parameter Hessian. It then defines a consumer-side group that captures the functional perturbation of removing its associated full dependency group, selects groups globally under a physical parameter budget, and applies ridge reconstruction after selection. We validate COSP through comprehensive experiments: (1) verifying its second-order derivative propagation and consumer-side saliency, (2) comparing with established pruning baselines, and (3) ablation studies on key design choices and computational cost. COSP achieves the best mean decoded quality for every evaluated modality and parameter reduction, and ranks first in 125 of 132 matched evaluation settings.

Primary area:
optimization
Keywords:
Implicit Neural Representations, Structured Pruning, Second-order Optimization, Neural Network Compression, Post-training Pruning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zTgtOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

COSP studies post-training structured pruning of implicit neural representations. It combines consumer-side scoring, diagonal curvature, global budgeted selection, and ridge reconstruction to produce compact dense INR networks. Experiments evaluate reconstruction quality and deployment speedup across images, biomedical volumes, SDFs, and videos.

Critical strengths
  • The motivation of this work is clear, and the main steps of the compacting pipeline are reasonably easy to follow.

  • The experiments cover several signal modalities, include component ablations, and report deployment speedups.

Critical weaknesses
  1. The methodological contribution may need clearer positioning. The summary of contribution (p. 2, lines 80–82) indicates that module-wise second-order propagation is a main contribution but Eq. (8) seems to be similar with the Hessian composition rule used in existing modular Hessian backpropagation [a]. This related works is not covered. From my part, the contribution primarily appears to be an application to INR pruning, and the authors should distinguish its pruning-specific advance from the established derivative machinery.

I think this submission is more like an application-oriented manuscript, while find that authors choose Primary Area: optimization.

  1. The empirical evidence seems not to be sufficient, especially on the baseline comparison. The baselines listed in lines 306–313 mainly cover conventional pruning criteria (The latest one is empirical Fisher (Liu et al., 2021)). Table 3 usefully controls for ridge reconstruction, but only at 50% reduction. Under this control, COSP reaches 21.7 dB on Video versus 21.5 for Random, and 0.55 F1 on SDF versus 0.53 for Fisher. These smaller margins make stronger comparisons particularly important.

[a] Modular Block-diagonal Curvature Approximations for Feedforward Architectures – aistats 2020 Eq. (7)

Questions for authors
  1. What exact discrepancy and targets are used for calibration? Eq. (1) compares the pruned model with the dense model (p. 2, lines 101–104), while all methods reportedly share the calibration objective (p. 6, lines 306–307). For squared output discrepancy, the per-sample gradients vanish at the dense checkpoint. How are non-degenerate Taylor and empirical-Fisher scores obtained?
  2. How were the baselines selected, and what establishes their competitiveness for this setting?
5 comments
AuthorsOpenReview ↗

Response to Reviewer zTgt: Calibration Targets and Pruning Contributions Beyond HBP

Thank you for recognizing the clear motivation, compression pipeline, multimodal evaluation, and deployment benefits. We have clarified the calibration targets, positioned the pruning contribution relative to HBP, and expanded the comparisons and controlled experiments. Revision highlights are blue. This reply addresses the scoring objective and methodological contribution. The second addresses competitiveness and reconstruction.

Q1. Calibration targets and nonzero Taylor/Fisher scores

Pruning scores use MSE against data labels, not the dense model's own outputs. Images and videos use observed pixels, volumes use observed intensities, and SDFs use supervision constructed from surface points and local geometry. The loss averages squared errors over coordinates and output channels in each model's output encoding. We have added a specific clarification of Eq. (1) in the main paper.

Your zero-gradient argument applies when the target is the same frozen model's output. The main comparison instead uses task labels, so the scoring signal comes from task residuals. We checked both target choices on all 33 original checkpoints, keeping each model and its calibration coordinates fixed:

  • With task labels, the checked Taylor and Fisher scores are nonzero on all 33 checkpoints.
  • With each model's own outputs as labels, the residuals, gradients, and corresponding scores are zero.
  • In 528 paired FP64 single-sample checks, every task-label gradient is nonzero and every corresponding self-output-label gradient is zero.
  • The 48 original Taylor/Fisher score arrays saved for the image experiments also contain no exactly zero entries.

Thus, the nonzero scores are supported by both the saved image results and the paired diagnostics, rather than an appeal to numerical noise.

Scoring and reconstruction use different targets. Compact affine layers are fitted to dense pre-activations, and updates are accepted using output discrepancy from the dense model on the calibration set. The main paper's Ridge Reconstruction section now states this criterion. Supplement A, under Method Overview and Notation, and Supplement B, under Frozen shared protocol, explain the target roles and scoring checks.

W1. Relationship to HBP and the pruning contribution

We have cited Dangel et al. (2020) in the Introduction and clarified that COSP uses modular second-order propagation. The contribution is how curvature is converted into deletion scores and structure decisions aligned with physical compaction.

HBP provides a modular mechanism for curvature computation. COSP addresses the further pruning decisions: which perturbation represents deletion, how its score relates to the physically removed group, and how groups are selected under an actual compact parameter budget.

  1. Deletion model. COSP scores zeroing the consumer column. This has the functional effect of removing the neuron, while export removes the producer row, bias, and consumer column together.
  2. Scoring. Shared first-order and propagated diagonal-curvature states evaluate candidate consumer perturbations without constructing a global parameter Hessian.
  3. Budget and recovery. Selection counts affine weights and biases in the exported network. Reconstruction then fits the surviving affine mappings with the structure fixed.

These pruning decisions are not specified by the HBP propagation rule itself. Their value is tested directly. In the single-neuron experiment, consumer-side Diag has a Spearman correlation of 0.9825 with absolute damage, compared with 0.8009 for the full-dependency estimate.

A controlled ablation fixes consumer grouping, calibration, selection constraints, and ridge, changing only whether diagonal curvature is added to the first-order score:

Kodak24 reductionMean PSNR gainImages improved
10%+2.888 dB24/24
30%+3.386 dB24/24
50%+2.059 dB21/24

Sections 3.1 and 3.3 and supplementary Table 10 report this evidence. It demonstrates practical value from the deletion model and curvature information beyond the shared reconstruction.

Positioning within optimization

We selected optimization because COSP studies structure selection in a frozen network under a physical parameter budget, using local second-order damage estimates. INRs provide the evaluation setting, while the method concerns deletion modeling, budget-constrained selection, and recovery. The revised Introduction makes this pruning-method contribution explicit.

AuthorsOpenReview ↗

Response to Reviewer zTgt: Recent Baselines and Evidence Beyond Reconstruction

This reply addresses the requested stronger comparisons and the interpretation of the reconstruction control.

Q2. Baseline selection and competitiveness

The original comparison covers classical first-order, second-order, and other pruning criteria. To test competitiveness against recent structured-pruning approaches, we added DepGraph, Isomorphic, FLAP, and GISP across all 33 original checkpoints and four physical parameter budgets. Methods use the same dense models and calibration samples within each task, with matched parameter reduction after compact export. The coordinate-MLP adaptations and selection constraints are described in Supplement B.

Among the nine methods, COSP achieves the highest mean primary quality in all 16 task-budget settings and ranks first in 119/132 individual settings. Primary quality is PSNR for images, volumes, and videos, and Global F1 for SDFs. Table 1 reports the following margins over the strongest baseline:

Task30% reduction50% reduction
Images, PSNR+5.14 dB+5.11 dB
Volumes, PSNR+4.60 dB+8.63 dB
Videos, PSNR+3.86 dB+3.54 dB
SDFs, Global F1+0.06+0.22

Figure 3 and supplementary Tables 3–6 now report the expanded comparisons, and Figures 4–5 have been updated accordingly.

We also expanded the primary video comparison from three clips to seven by adding Beauty, Bosphorus, ReadySetGo, and YachtRide. This compares eleven methods at four budgets, using the first 32 frames of each clip at 256×512 resolution. At the two central budgets:

ReductionCOSP mean PSNRStrongest baselineMargin
30%25.06521.372+3.693
50%21.83418.409+3.425

COSP achieves the highest mean PSNR at all four budgets. Supplementary Tables 15–18 provide dense quality, complete PSNR/SSIM comparisons, and per-clip results. These experiments strengthen the evidence beyond the original classical baselines and video sample.

W2. Video and SDF margins under common reconstruction

Table 3 reflects the complementary roles of structure selection and parameter recovery. COSP selects a structure with higher video quality before recovery, while reconstruction can give other selections larger improvements and narrow the final gap. We therefore use separate controls to quantify selection and curvature contributions.

Supplementary Table 12 adds an independent video selection control at 10/30/50/70% reduction with common reconstruction. At 10%, 30%, and 50%, COSP's mean PSNR advantages over the strongest baseline are 6.848, 1.693, and 0.258 dB. The complete budget results are reported under Selection and reconstruction across budgets.

Supplementary Table 10 isolates curvature with consumer grouping, selection constraints, and reconstruction fixed. In addition to the Kodak24 gains in our first reply, SDF mean Global F1 improves by 0.0382 and 0.0093 at 30% and 50% reduction. Video mean PSNR improves by 0.828 dB at 30%. These are increments from adding curvature under matched conditions, distinct from the complete-method margins in Table 1.

We further tested multiple-group selection before reconstruction. On Kodak24 at 30%/50% reduction, adding curvature raises mean signed Spearman correlation from 0.076/0.061 to 0.697/0.396. COSP yields lower actual task MSE than first-order selection at the same layer widths on 24/24 and 23/24 images. Supplementary Tables 13–14 provide the complete diagnostics. This gives additional evidence that the scoring design has selection value without relying on reconstruction.

Together, the clarified targets, explicit pruning contribution, recent-method comparisons, and controlled gains address the methodological and empirical concerns. We hope this evidence supports a favorable reassessment.

AuthorsOpenReview ↗

Follow-up to Reviewer zTgt on Calibration, Methodological Contributions, and Comparisons

Thank you again for your thoughtful feedback. We have clarified the distinction between scoring and reconstruction targets, with paired checks across all original checkpoints confirming nonzero Taylor/Fisher scores under the task targets. We have also cited HBP and distinguished COSP’s pruning-specific contributions from the established derivative machinery. To strengthen the empirical evidence, we added comparisons with four recent pruning methods, shared-reconstruction controls across all four budgets, and experiments isolating the contribution of curvature. Revisions are highlighted in blue.

We would appreciate your assessment of whether these clarifications and additional results address your concerns, particularly regarding the calibration objective, the contribution relative to HBP, and the value of scoring beyond reconstruction. We would be happy to clarify any remaining questions.

Thank you for your time and consideration.

Reviewer_SE1bOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

COSP proposes a post-training, one-shot structured-pruning method for implicit neural representations (INRs). It scores neuron removal using consumer-side functional perturbations and module-wise second-order derivative propagation with a low-cost diagonal approximation. Candidate neurons are ranked globally under an actual parameter budget, and fixed-mask ridge reconstruction is applied after physical compaction. Experiments on image, biomedical-volume, SDF, and video SIRENs report stronger reconstruction quality than six classical pruning baselines at 10–70% parameter reduction.

Critical strengths

The paper addresses a practical problem: physically reducing the size and inference cost of pretrained coordinate-based INRs while preserving reconstruction fidelity. The distinction between consumer-side scoring and full physical neuron removal is clearly motivated. The experimental study spans four signal modalities, several pruning ratios, and multiple classical pruning baselines. Derivative-consistency checks, scoring-group comparisons, curvature ablations, ridge-reconstruction controls, and runtime/memory measurements are valuable components of the evaluation.

Critical weaknesses

The main comparisons are restricted to Random, Magnitude, Taylor, Fisher, OBD, and OBS. While these cover classical scoring families, they do not establish superiority over recent, competitive structured-pruning approaches applicable to coordinate MLPs. The paper also does not provide a matched end-to-end rate–distortion comparison with INR compression systems; differences in representation and coding assumptions should be acknowledged when designing such comparisons. The contribution of the proposed saliency score is not fully isolated from ridge reconstruction. In the shared-ridge control at 50% parameter reduction (Table 3), video PSNR is 21.7 dB for COSP versus 21.5 for Random, 21.4 for Taylor, and 21.3 for Fisher. This narrow margin contrasts with the larger gaps in the default comparison and raises questions about the independent benefit of the second-order scoring mechanism after reconstruction. Shared-reconstruction comparisons across all pruning ratios would be more informative. Generalization is insufficiently demonstrated. All main compression experiments use SIREN models, even though derivative correctness is checked with other activations. The video evaluation includes only three UVG clips, and SDF evaluation includes only two scenes. The strong single-neuron score correlations do not by themselves establish the quality of one-shot ranking when many neurons are removed simultaneously.

Questions for authors
  1. Could the authors provide direct, protocol-matched comparisons against the closest applicable recent methods?
  2. Why does the advantage over simple baselines become small on video after identical ridge reconstruction ? Can the authors report shared-ridge results at every pruning ratio and quantify the incremental gain attributable to saliency selection alone?
  3. Does COSP retain its gains on non-SIREN INR architectures and a larger, more diverse video benchmark? How well do individual-group Taylor scores predict the actual cumulative distortion at aggressive pruning ratios, where cross-group interactions are omitted?
3 comments
AuthorsOpenReview ↗

Response to Reviewer SE1b: Baseline Comparisons, Reconstruction Controls, and Compression Scope

Thank you for recognizing COSP's deployment value, the distinction between consumer-side scoring and physical deletion, and the derivative checks, ablations, and cost measurements. We have expanded the comparisons and validation in the revised paper and supplement, with revision highlights in blue. This reply addresses recent methods, reconstruction, and codec scope. The second addresses generalization and simultaneous deletions.

Q1. Direct comparisons with recent methods

We added DepGraph, Isomorphic, FLAP, and GISP across the original 33 checkpoints and four physical parameter budgets. Methods use the same dense models and calibration samples within each task, with matched parameter reduction after compact export. Supplement B documents the coordinate-MLP adaptations and selection constraints.

In the expanded nine-method comparison, COSP achieves the highest mean primary quality in all 16 task-budget settings and ranks first in 119/132 individual settings. Primary quality is PSNR for images, volumes, and videos, and Global F1 for SDFs. Table 1 reports margins over the strongest baseline:

Task30% reduction50% reduction
Images, PSNR+5.14 dB+5.11 dB
Volumes, PSNR+4.60 dB+8.63 dB
Videos, PSNR+3.86 dB+3.54 dB
SDFs, Global F1+0.06+0.22

Figure 3 and supplementary Tables 3–6 now include the expanded comparison at all budgets. These results directly strengthen the evidence for COSP's competitiveness as a complete compression method.

Q2. Why the video margin narrows after reconstruction

Table 3 reflects the distinct roles of structure selection and parameter recovery: COSP selects a structure with higher quality before recovery, while reconstruction can yield larger improvements for other selections, narrowing the final gap. We therefore quantify the additional value of selection and curvature separately through controls across all budgets with common reconstruction and a curvature ablation with fixed grouping.

Following your suggestion, supplementary Table 12 adds an independent video selection control at all four reductions, 10/30/50/70%, under common reconstruction. COSP's mean PSNR advantages over the strongest baseline are 6.848, 1.693, and 0.258 dB at 10%, 30%, and 50%, respectively. The complete budget results are reported under Selection and reconstruction across budgets. This control quantifies the quality differences remaining between selected structures after the same recovery procedure.

To isolate curvature itself, supplementary Table 10 fixes consumer grouping, calibration, selection constraints, and ridge, changing only whether diagonal curvature is added to the first-order score:

Kodak24 reductionMean PSNR gain from curvatureImages improved
10%+2.888 dB24/24
30%+3.386 dB24/24
50%+2.059 dB21/24

The corresponding video gain at 30% reduction is 0.828 dB. Table 10 retains all four budgets and modalities. The two controls distinguish the contribution of structure selection from the incremental value of curvature, rather than attributing every post-reconstruction difference to the second-order term.

Comparison scope with INR codecs

The Introduction clarifies that COSP compacts an existing trained INR to reduce stored parameters and inference cost. Supplement B explains that a matched end-to-end rate–distortion comparison with codecs requires the same data, resolution, and distortion metric, with rate including the complete bitstream and decoding metadata.

Our measured deployment benefit is 1.27–1.94× speedup in forward inference and full-signal decoding at 30%/50% parameter reduction (Table 1). This supports the practical value of structural compaction. The present experiments measure compact parameters and inference cost, not matched entropy-coded bitstream rate–distortion performance.

AuthorsOpenReview ↗

Response to Reviewer SE1b: Generalization and Multiple-Group Deletion

This reply addresses your questions about generalization beyond SIREN, broader video coverage, and the predictive value of scores when multiple groups are deleted.

Q3(a). Compression beyond SIREN

Supplementary Table 8 adds actual compression experiments on Fourier-feature MLPs with Softplus and Tanh activations, comparing with seven other algorithms. The seven Softplus and six Tanh checkpoints were selected using dense quality thresholds before pruning. At 30% and 50% physical parameter reduction, mean PSNR is:

ArchitectureReductionCOSP (dB)Strongest baseline (dB)Margin (dB)
Softplus30%24.54620.556+3.990
Softplus50%22.16118.358+3.803
Tanh30%22.99922.004+0.995
Tanh50%20.07319.725+0.348

COSP leads all seven baselines in these four settings. These are actual compact-network quality results, extending the evidence beyond derivative correctness checks with other activations.

Q3(b). Broader video evaluation

We expanded the primary video analysis from three clips to seven, adding Beauty, Bosphorus, ReadySetGo, and YachtRide to CityAlley, Jockey, and HoneyBee. We compare eleven methods at four budgets. Each clip uses its first 32 frames at 256×512 resolution, with the same dense model and 65,536 calibration coordinates across methods within that clip.

ReductionCOSP mean PSNRStrongest baselineMargin
10%34.48925.707+8.782
30%25.06521.372+3.693
50%21.83418.409+3.425
70%20.16715.404+4.763

COSP achieves the highest mean PSNR at all four budgets. Supplementary Tables 15–18 report dense quality, complete PSNR/SSIM comparisons, and per-clip results.

We also trained a SIREN on Dragon and completed an eleven-method pruning comparison at four physical parameter budgets using the final 200,000-step checkpoint. Dense Global F1 is 0.6058. All methods use the same model, 8,192 calibration coordinates, and surface evaluation protocol:

ReductionCOSP F1Strongest baseline F1Margin
10%0.59630.5967−0.0003
30%0.57100.5546+0.0164
50%0.53470.3766+0.1581
70%0.46340.1911+0.2724

COSP achieves the highest F1 at 30%, 50%, and 70% reduction, extending the evidence for compression gains beyond the original two scenes. Supplementary Table 19 reports the complete comparison.

Q3(c). Simultaneous deletion of multiple groups

We added Multiple-group deletion to Supplement B, evaluating all 33 original checkpoints beyond single-neuron perturbations. Each model uses 2,048 scoring coordinates and 2,048 disjoint evaluation coordinates from the original calibration pool. Candidate masks are constructed without using the evaluation split.

Within each model and budget, candidates have the same layer widths and therefore the same compact parameter count. We compare signed cumulative predictions with actual task MSE changes before reconstruction, deduplicate identical masks, and compute correlations within each candidate pool before averaging over models.

On Kodak24, both 30% and 50% reduction involve deletions across layers:

ReductionFirst-order SpearmanWith diagonal curvatureLower MSE than first-order selection
30%0.0760.69724/24 images
50%0.0610.39623/24 images

COSP also yields lower MSE than the random-candidate mean on all 24 images at both budgets. Thus, the evidence includes actual selection gains, not only score correlations, and does not depend on reconstruction.

A sensitivity check restricted to random candidates gives the same direction of improvement: correlations rise from 0.125/0.037 to 0.411/0.253 at 30%/50% reduction. Supplementary Tables 13–14 report ranking performance and cumulative distortion prediction errors across all budgets and modalities. These results support gains in ranking and selecting multiple groups at 30% and 50% reduction.

We hope the direct recent-method comparisons, separate attribution controls, and expanded generalization evidence address your concerns and strengthen your assessment of COSP.

AuthorsOpenReview ↗

Follow-up to Reviewer SE1b on Additional Experiments and Clarifications

Thank you again for your detailed feedback. Following your suggestions, we added comparisons with four recent pruning methods, shared-reconstruction controls at all four pruning ratios, and an ablation isolating the contribution of curvature. We also extended the evaluation to non-SIREN architectures, seven video clips, and the Dragon SDF scene, and tested multiple-group selection before reconstruction. The revised manuscript clarifies the scope of comparison with INR codecs, with revisions highlighted in blue.

We would appreciate your assessment of whether this additional evidence addresses your concerns, particularly regarding the contribution of scoring beyond reconstruction and generalization beyond the original models and scenes. We would be happy to clarify any remaining questions.

Thank you for your time and consideration.

Reviewer_8LSROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes COSP, a one shot method to reduce the size of trained implicit neural representation networks. The authors does this by removing hidden neurons and exporting a smaller network. They first runs a single backward pass that propagates both gradients and a diagonal curvature approximation through the frozen network. It then scores each neuron according to the damage estimate from deleting its consumer side weights. They remove neurons until a parameter budget is met. Finally, they refit the remaining layers with ridge regression to repair the damage. They evaluate the proposed methods compression performance on images, videos, biomedical volumes and SDFs.

Critical strengths

Paper presents a new way of compacting implicit neural representation networks and the method is cheap and fast while achieving consistent gains with real parameter compaction across their tested data modalities. Furthermore, it leads on all four modalities when every method gets the same ridge repair.

Each approximation is validated against exact computation: Full curvature propagation matches nested autograd to numerical precision and the diagonal variant closely tracks the exact Hessian diagonal.

Critical weaknesses

Novelty is modest and I would argue that and closely related prior work is either uncited or not positioned properly. For example, Kwon et al. (2022), which the paper cites only as a generic second-order method already does the post-training layer-wise least-squares reconstruction. It would be better if the authors can state how their method is different from such work and why that difference matters.

The baselines are all classic pruning methods. Authors need to compare their model with new methods.

Questions for authors

The paper never reports the dense models own quality. It would be good to report them for comparison.

Authors only use SIREN models. It would be good to test this method against other INR models.

Table 2 presents image degradation to be identical (18.559 dB) for the Diagonal, Exact-Diagonal and Exact-Full variants. Can you clarify if this is a typo or not? I observe the same repeated pattern in video degradation.

I also noticed a typo in refereeing to equations("Eq. equation").

3 comments
AuthorsOpenReview ↗

Response to Reviewer 8LSR: Pruning Contributions and Recent Baselines

Thank you for recognizing COSP's efficiency, physical compaction, and derivative validation. We have revised the main paper and supplement, with revision highlights in blue. This reply addresses the contribution and recent-method comparisons. The second reply addresses dense quality, other INR architectures, and reporting.

W1. Distinction from Kwon et al. and why it matters

We agree that post-training least-squares reconstruction has precedents. COSP's contribution is the combination of consumer-side deletion scoring, selection under actual compact parameter budgets, and reconstruction of the surviving affine mappings. The Introduction and supplementary Table 9 now distinguish these components:

AspectKwon et al. (2022)COSP
ScoreFisher approximation in head/filter mask spaceFirst-order and propagated diagonal curvature for consumer-column deletion
SelectionResource-constrained mask search and intra-layer rearrangementOne global ranking from the frozen INR
BudgetFLOPs or hardware latencyActual affine weights and biases after compact export
ReconstructionTune retained mask coefficients to match dense layer outputsFit compact affine weights and biases to dense pre-activations

The distinction matters because scoring and physical deletion need not use the same parameter group. Zeroing a consumer column has the functional effect of deleting its neuron. Export then removes the producer row, bias, and consumer column together. COSP thus evaluates the consumer perturbation while counting the full physical saving. In the original single-neuron validation, consumer-side Diag achieves a Spearman correlation of 0.9825 with absolute deletion damage, versus 0.8009 for the full-dependency estimate.

We also isolate the contribution beyond reconstruction. With consumer grouping, calibration, selection constraints, and ridge fixed, adding diagonal curvature to the first-order score gives:

Parameter reductionKodak24 mean PSNR gainImages improved
10%+2.89 dB24/24
30%+3.39 dB24/24
50%+2.06 dB21/24

Section 3.3 reports these gains, and supplementary Table 10 contains the complete four-budget control. The matched reconstruction makes this evidence specific to the additional curvature information, rather than attributing all quality recovery to ridge.

The new multiple-group diagnostic further tests the design before reconstruction. At 30%/50% reduction on Kodak24, diagonal curvature raises mean signed Spearman correlation from 0.076/0.061 to 0.697/0.396. COSP selects masks with lower actual task MSE than first-order selection at the same layer widths on 24/24 and 23/24 images. Supplementary Tables 13–14 provide the full diagnostics. Together, these experiments support both the deletion model and its practical selection value.

W2. Comparisons with recent methods

We added DepGraph, Isomorphic, FLAP, and GISP across all 33 original checkpoints and four physical parameter budgets. Methods use the same dense models and calibration samples within each task, with matched parameter reduction after compact export. The coordinate-MLP adaptations and selection constraints are documented in Supplement B.

In the expanded nine-method comparison, COSP achieves the highest mean primary quality in all 16 task-budget settings and ranks first in 119/132 individual settings. Primary quality is PSNR for images, volumes, and videos, and Global F1 for SDFs. Table 1 gives the following margins over the strongest baseline:

Task30% reduction50% reduction
Images, PSNR+5.14 dB+5.11 dB
Volumes, PSNR+4.60 dB+8.63 dB
Videos, PSNR+3.86 dB+3.54 dB
SDFs, Global F1+0.06+0.22

Figure 3 and supplementary Tables 3–6 now include these methods and all budgets. Figures 4–5 also show updated qualitative or distributional comparisons.

We additionally expanded the primary video analysis from three to seven clips by adding Beauty, Bosphorus, ReadySetGo, and YachtRide. This comparison includes eleven methods at four budgets. Each clip uses its first 32 frames at 256×512 resolution. Mean PSNR is:

ReductionCOSPStrongest baselineMargin
10%34.48925.707+8.782
30%25.06521.372+3.693
50%21.83418.409+3.425
70%20.16715.404+4.763

COSP leads at all four budgets. Supplementary Tables 15–18 report dense quality, PSNR/SSIM, and per-clip results, plus an eight-clip analysis including ShakeNDry. These additional results strengthen the evidence beyond the original video sample without changing the original 119/132 comparison.

AuthorsOpenReview ↗

Response to Reviewer 8LSR: Dense Quality, Generalization, and Reporting Clarifications

Thank you for the questions on dense quality, generalization, and repeated values. Below are the corresponding additions and verification results.

Q1. Quality of the dense models

We added dense PSNR labels to Figures 4 and 7 and dense Global F1 to the caption of Figure 6. Supplementary Table 7 reports per-instance quality for all 33 dense models in the main comparison, allowing comparison before and after compression.

Q2. Compression beyond SIREN

We added actual compression experiments on Fourier-feature MLPs with Softplus and Tanh activations, covering seven and six checkpoints, respectively, and comparing with seven other algorithms. Mean PSNR at 30%/50% parameter reduction is shown below. Supplementary Table 8 provides the complete comparison.

ArchitectureReductionCOSP (dB)Strongest baseline (dB)Margin (dB)
Softplus30%24.54620.556+3.990
Softplus50%22.16118.358+3.803
Tanh30%22.99922.004+0.995
Tanh50%20.07319.725+0.348

COSP leads all seven baselines in these four settings, supporting its practical compression competitiveness on non-sinusoidal INRs.

Q3. Repeated values in Table 2

We verified the original records: the repeated values are not typos. Table 2 reports medians. The three image variants share the middle records and have identical masks for those records. For video, Diagonal and Exact-Diagonal both take their median from HoneyBee at 30% reduction. Equal medians do not imply identical results throughout the evaluation. Supplementary Table 11 gives the means, medians, and source instances. The original table values are unchanged.

Q4. Equation-reference formatting

We corrected the duplicated “Eq. equation” prefix in both PDFs and checked the equation references.

We hope the clarified contribution, expanded comparisons, and additional validation in these two replies address your concerns and support a favorable reassessment.

AuthorsOpenReview ↗

Follow-up to Reviewer 8LSR on Revisions and Additional Evidence

Thank you again for your thoughtful feedback. We have clarified COSP’s distinction from Kwon et al. and added controlled experiments isolating the contribution of curvature beyond reconstruction. We also added comparisons with four recent pruning methods, compression experiments beyond SIREN, and dense-model quality metrics. The repeated values in Table 2 are explained in the supplement, and the equation-reference formatting has been corrected. Revisions are highlighted in blue.

We would appreciate your assessment of whether these revisions address your concerns, particularly regarding the pruning-specific contribution and competitiveness against recent methods. We would be happy to clarify any remaining questions.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.