- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper replaces the fixed B-spline basis in SplineCNN with learnable rational basis functions. Safe denominators, Chebyshev representations, spline/PCA initialization, and variance calibration support training. The convolution retains a bank of feature-transform matrices while allowing a compact multivariate basis. Experiments across geometric matching, event recognition, superpixel classification, and point-cloud segmentation suggest parameter-efficiency benefits and some accuracy improvements.
Critical strengths
- Practical parameter efficiency. The method addresses the growth of the matrix bank with the size of a tensor-product spline grid. In higher-dimensional settings, compact bases obtain performance close to much larger spline models. This is a useful practical contribution even where accuracy gains are small.
- Breadth of validation and implementation. Replacing convolution layers in several tasks and host architectures helps test whether the approach generalizes beyond a single pipeline. The structured initialization, numerical representation, and supplied configurations and logs strengthen the implementation contribution.
Critical weaknesses
- Selection using test performance. Table 1 and the supplementary documentation indicate that PascalVOC/DGMC and the NMT columns report the best test epoch, whereas SPair-71k/DGMC uses the best validation epoch. Taking the maximum over test scores makes the test set part of result selection and introduces optimistic bias. The same procedure for baselines does not remove this concern. The issue is the epoch-selection rule, not the absence of patience-based early stopping. All affected comparisons should use validation-selected checkpoints or a final epoch fixed independently of test outcomes. The supplement retains a final-epoch advantage in PascalVOC/DGMC and reports small best-test versus final-epoch differences for NMT, but a consistent corrected evaluation is still needed before the headline gains can support acceptance.
- The rational form is not sufficiently isolated. The input domain remains . In , reducing reduces the matrix parameters; replacing splines at the same does not do so and adds function coefficients. For smaller , the initialization first compresses the spline space and then fits rational functions to the retained components. The contribution of compression, initialization, adaptation, and rational parameterization therefore needs separation. MLP and frozen-rational controls are relevant, but do not replace a direct PCA-compressed spline baseline with matched total capacity and comparable initialization. This distinction determines whether rational functions themselves justify the main novelty claim.
Questions for authors
- Clarification and decisive evidence: Which conclusions remain when every affected method uses a test-independent epoch-selection rule?
- Decisive comparison: How does the method compare with directly retaining the same PCA spline components, and with learned bases matched in total capacity and initialization quality?
- Clarification: How should the general inference-speed claim be reconciled with the , 32-to-32-channel microbenchmark, where the rational forward pass is slower than PyG SplineConv? Please also account for initialization cost.
3 comments
Q3. Speed claim and initialization cost.
The reviewer is right that the claim was too general. We reran the benchmark (method section, Appendix D). The slower forward pass comes from a variant of our Triton kernel that uses a memory-saving aggregation order. Our RBF-GNN triton kernel can evaluate a layer in two orders with identical outputs, which we have not mentioned in the initial submission:
- Weight-first forms each message on its edge. It is fast but stores values, since all rational functions are non-zero on every edge, whereas PyG evaluates only the non-zero B-splines.
- Basis-first first sums over the neighbours and then applies the . It needs about PyG's memory, but it passes over the edges times, so its forward pass is slower.
We give below a micro benchmark with both orders. We use one layer on a synthetic graph with 50,000 nodes, edges and 32 → 32 channels (RTX 5090, degrees (10,8), Triton basis kernel), on the pseudo-coordinate grids of our three benchmarks. Each cell: forward (ms) / forward + backward (ms) / peak memory (GB). SplineCNN has = 25, 125 and 8.
| Layer | 2-D, (keypoints) | 3-D, (FAUST) | 3-D, (event cameras) | |
|---|---|---|---|---|
| SplineCNN (PyG, fused) | 2.1 / 14.7 / 0.40 | 4.7 / 24.5 / 0.44 | 3.5 / 32.4 / 0.44 | |
| RBF-GNN, weight-first | 4 | 1.9 / 4.7 / 1.10 | 2.2 / 5.3 / 1.10 | 2.2 / 5.3 / 1.10 |
| RBF-GNN, weight-first | 8 | 2.3 / 6.6 / 2.08 | 2.7 / 7.8 / 2.08 | 2.7 / 7.9 / 2.08 |
| RBF-GNN, weight-first | 16 | 4.2 / 12.6 / 4.05 | 5.2 / 15.1 / 4.05 | 5.2 / 15.5 / 4.05 |
| RBF-GNN, weight-first | 32 | 8.3 / 25.0 / 7.98 | 10.1 / 30.1 / 7.98 | 10.3 / 31.1 / 7.98 |
| RBF-GNN, basis-first | 4 | 3.6 / 9.2 / 0.42 | 3.8 / 9.8 / 0.42 | 3.8 / 9.8 / 0.42 |
| RBF-GNN, basis-first | 8 | 7.1 / 18.1 / 0.47 | 7.6 / 19.3 / 0.47 | 7.6 / 19.3 / 0.47 |
| RBF-GNN, basis-first | 16 | 14.3 / 36.0 / 0.58 | 15.3 / 38.5 / 0.58 | 15.3 / 38.5 / 0.58 |
| RBF-GNN, basis-first | 32 | 28.7 / 74.0 / 0.79 | 30.5 / 79.2 / 0.79 | 30.5 / 79.2 / 0.79 |
- The speed advantage depends on the aggregation order, basis size and channel width. Weight-first can outperform PyG at higher memory cost.
- Initialization cost: We only pay a one-time CPU fit that does not depend on the data and is cached and shared by all layers with the same configuration. It takes 2 to 3 s in 2-D and 36 to 41 s in 3-D, but about 31 minutes on the 5-D grid of our CIFAR-10 experiment. Usually, initialization is a negligible share of training: under 1% in 2-D and about 0.2–0.4% on the event cameras. On FAUST it is about a quarter of a run, because those runs take only about 3 minutes.
Thanks for the clarification. I have raised my score.
We thank the reviewer for the constructive review. We address reviewer's concerns below:
We agree that the selection should have been done on validation data. We have therefore retrained all affected results, baselines included, and now report the test accuracy at the epoch with the best validation accuracy:
MNIST, CIFAR-10 and ShapeNet-Part report the final epoch of a fixed schedule and are unchanged. Under DGMC, every method family also gets the same learning-rate grid with five seeds per rate, selected on mean validation accuracy (general response, point 2).
Overall the selection bias was small. Crucially, the relative ordering remains the same.
We report the best variant of each method, as submitted and after retraining. The revised variants are selected on validation but for DGMC also come after a more exhaustive learning rate grid search, which resulted in higher DGMC scores overall. We give the difference at the best-validation epoch and, as the test-independent alternative the reviewer names, at the final epoch (Welch's t-test). Overall, the absolute numbers have changed slightly but relative ordering stayed mostly the same.
Note that using the new selection rule changed also the training set, since for PascalVOC we took out 10% for validation and for FAUST we removed subject 7. RBF-GNN uses degrees (10,8).
To disentangle compression, initialization, adaptation, and rational parameterization we have conducted experiments in Table 7 (Appendix B).
Every row has the same K weight matrices, and total parameters differ by at most 0.11% (at K=9 under DGMC: 3.741M rational, 3.738M PCA-spline, 3.740M MLP, 3.737M Gaussian). All rows use the published learning rate:
*: On FAUST, SplineCNN with small K degrades heavily but works well with higher K, see Table 2.
Conclusion. The rational form is the best form at equal initialization. Against the exact PCA-compressed splines, however, the margin is small, yet significant in three of seven settings. Against default-initialized learned bases, a significant part of the gain comes from the initialization, with the rational form adding some more performance on average.