- 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
- 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.
- 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
- 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?
- How were the baselines selected, and what establishes their competitiveness for this setting?
5 comments
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:
| Task | 30% reduction | 50% 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:
| Reduction | COSP mean PSNR | Strongest baseline | Margin |
|---|---|---|---|
| 30% | 25.065 | 21.372 | +3.693 |
| 50% | 21.834 | 18.409 | +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.
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.
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:
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.
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:
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.