acceptodds
Under review as a conference paper at ICLR 2027

WaveKernel: Continuous Kernel Synthesis via Optical Wave Propagation

Abstract

Continuous-time models aim to learn discretization-consistent signal representations, with their expressive properties significantly influenced by the functional parameterization of local temporal operators. Prior approaches largely rely on predefined basis functions or neural networks, leaving the effective parameterization of continuous kernels an important open challenge. We propose WaveKernel, a continuous convolution operator that synthesizes kernels by evaluating the optical field generated by a slab-type diffractive model at continuous spatial positions. Learnable amplitude–phase modulation units jointly shape the spatial profile of the optical field, thereby defining a continuous kernel expanded in a basis of optical impulse responses. By introducing a structured inductive bias derived from optical wave propagation and mapping relative temporal coordinates onto continuous spatial positions for kernel evaluation, the proposed parameterization allows the same learned coefficients to be evaluated across different sampling rates. Integrated into a shared time-series architecture, WaveKernel improves seven-rate mean accuracy over spline-based continuous convolution on three sleep-staging benchmarks, while maintaining stable performance across the evaluated sampling rates.

Primary area:
learning on time series and dynamical systems
Keywords:
continuous-time models, continuous convolution, neural operators, kernel learning, photonic computing, diffractive optics, physiological time series, sampling-rate robustness

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_J1mjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes WaveKernel, a continuous-kernel parameterization that synthesizes kernels as learned combinations of fixed optical propagation responses. A fixed optical excitation propagates through a simulated slab waveguide to a set of latent positions, following the scalar Huygens–Fresnel principle. Mode-wise complex coefficients provide task-specific adaptation, and a second passive propagation produces a coordinate-dependent complex optical response, whose real part gives the kernel. Averaged over seven sampling rates, WaveKernel outperforms spline by 1.57–3.66% on three sleep-staging datasets and trails it by 0.37% on ECG. Both kernels keep nearly constant accuracy across sampling rates. WaveKernel is less robust to strong broadband noise, reaching 42.08% accuracy at 10 dB white noise compared with 54.01% for spline.

Critical strengths
  • WaveKernel improves accuracy over the spline-based continuous kernel without losing cross-rate stability. It exceeds spline by 1.95, 1.57, and 3.66% on ISRUC, HMC, and SleepEOG (L321–L322) in all five folds (L969–L971) with a similar parameter count (338.15K vs. 319.59K), while its worst-case drop across seven rates (0.06–0.90%) stays comparable to spline (0.17–0.57%) and far below CKConv (10.53–36.68%) (Table 4).
  • The figures and tables are clear and well prepared.
Critical weaknesses

The specific challenge that WaveKernel addresses could be defined more clearly. Compared with the care given to the figures and tables, the line of argument in the text is less developed. The title and abstract present a general continuous-kernel parameterization, motivated by the statement that "the effective parameterization of continuous kernels remains an important open challenge" (L039–L040). A more concrete problem statement would help readers see why an optical basis is needed and how the experiments test it.

  • Problem formulation. The introduction notes that fixed bases confine kernels to prescribed function families, while neural kernels are sensitive to optimization (L045–L049). WaveKernel, however, is itself a finite expansion over a fixed basis, determined by diffraction physics and propagation geometry, with kernels spanning at most 2N2N dimensions per channel pair (L126–L129, L248–L252). It would help readers if the authors explained how this fixed basis avoids the limitation attributed to predefined bases, and which limitation the experiments are meant to test.
  • Choice of optics. Section 3.2 moves directly from the need for a learnable function family (L168) to slab waveguides (L191–L206). Since the model is purely digital (L464–L468), it would help to explain why propagation responses suit the problem better than other oscillatory bases, such as Fourier, Gabor, or chirp atoms.
  • Experimental domain. The introduction motivates the setting by physiological signals (L080–L083), but the only link between these signals and the basis is "a plausible inductive bias for physiological rhythms and transient patterns" (L431–L433), and consistent gains appear only on three sleep-staging datasets (L444–L446). For a general parameterization, more diverse benchmarks, such as those used for CKConv and FlexConv, would better show where the optical basis helps.

The comparison with neural kernels could be strengthened. The 10.6–23.5% margin over CKConv (Table 1) mainly reflects its collapse at unseen rates, which the authors relate to the 1/fs1/f_s scaling of its convolution and bias (L826–L827), an implementation choice rather than a property of MLP kernels. The 3.10–11.20% lead at the training rate (Tables 5–8) is measured against a single untuned CKConv with twice the parameters (L827; 677.25K vs. 338.15K). A tuned CKConv and FlexConv (L124–L126) at matched parameter counts would show whether the gain comes from the kernel family.

Questions for authors
  • Why are the responses of Eq. (2) a better basis for physiological kernels than other oscillatory bases with a comparable number of coefficients?
  • On the sleep datasets, all test rates are resampled from the same 35 Hz low-passed recording (L750, L755, L774, L845–L846), and U-Sleep, which simply resamples its input to the training rate (L847), is similarly stable (Table 4). Can the cross-rate experiment distinguish continuous kernels from input resampling? For example, how do the two compare when the model is trained on downsampled data and tested on the original high-rate recordings?
  • How does WaveKernel compare with a tuned CKConv (without the 1/fs1/f_s scaling of the bias) and with FlexConv at matched parameter counts, both at the training rate and across rates?
  • Which limitation of existing spline or neural kernels does WaveKernel resolve, and can it be measured directly? What do the learned kernels look like in time and frequency?
1 comment
AuthorsOpenReview ↗

Clarification of WaveKernel's Contribution and Experimental Evidence

We thank the reviewer for the constructive suggestions on the motivation and experimental evaluation of WaveKernel. We have refined the problem formulation and added a controlled comparison with alternative kernel bases.

1. Motivation and Propagation-Based Inductive Bias

WaveKernel investigates whether a compact, propagation-derived basis provides a useful inductive bias for continuous temporal kernels under a fixed parameter budget. Rather than claiming a strictly larger function space, our contribution lies in the optical propagation-based parameterization, which may enable more efficient representations of certain oscillatory kernels.

Optical propagation generates basis functions with coordinate-dependent amplitude envelopes and nonlinear phase variations. Unlike Fourier and standard Gabor functions with constant carrier frequencies, their local frequencies vary with the coordinate. Unlike quadratic-phase chirps, their phase curvature is generally nonconstant. Compared with spline basis functions, they explicitly incorporate oscillatory phase structures. These properties are further analyzed in Appendix A.5.

2. Controlled Comparison with Alternative Bases

To evaluate the effectiveness of the optical basis, we added a five-fold SleepEOG experiment comparing optical, spline, Fourier, Gabor, and chirp parameterizations under matched conditions, including identical architecture, sampling grids, normalization, training settings, and 337,129 trainable parameters.

Averaged over seven sampling rates, optical achieves 59.48 ± 1.58% accuracy, compared with 56.82 ± 1.38% for spline, 56.32 ± 1.99% for Fourier, 59.05 ± 1.26% for Gabor, and 58.94 ± 1.96% for chirp.

The optical basis retains improvements of 2.66 and 3.16 percentage points over spline and Fourier, respectively, while achieving comparable performance to Gabor and chirp. These results provide experimental support for the propagation-based parameterization under matched conditions. Details are provided in Section 4.1 and Appendix D (Table 9).

3. Cross-Rate Evaluation and Input Resampling

Our cross-rate experiments demonstrate that WaveKernel can directly evaluate learned continuous kernels at different sampling rates without retraining or changing the learned coefficients.

As the sleep recordings share a common 35 Hz low-pass-filtered source, input resampling to the training rate is also a valid approach to cross-rate evaluation. A comparison using low-rate training data and native high-rate test recordings would further distinguish these strategies. We have clarified the evaluation setting in Appendix B.4.

4. CKConv and FlexConv Comparisons

The CKConv baseline follows the publicly released NeurOCNN wrapper, which applies the 1/fs1/f_s factor after the affine convolution, scaling both the convolution output and bias. An alternative is to apply the scaling only to the convolution term while adjusting the bias to preserve the training-rate output.

Further comparisons with a bias-adjusted and tuned CKConv and a parameter-matched FlexConv would help assess the performance of WaveKernel against neural kernel parameterizations more comprehensively. We have clarified the CKConv implementation details in Appendix B.3.

5. Learned Kernels and Experimental Scope

Figure 6 illustrates the amplitude and phase characteristics of the optical basis functions. The learned kernels combine these responses through task-specific coefficients. A direct time- and frequency-domain analysis of the resulting learned kernels would further characterize their relationship to physiological rhythms.

We have refined the Introduction and Limitations to focus the current contribution on physiological time-series classification, supported by the results on three sleep-staging datasets and the new matched basis comparison.

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

This paper proposes WaveKernel for time-series classification. The main idea is to use optical propagation responses to build continuous convolution kernels. The model learns complex coefficients to combine these responses. The same learned coefficients can be used to evaluate the kernel at different sampling rates.

The authors replace the spline kernels in NeurOCNN with WaveKernel and keep the other main components the same. They evaluate the method on three sleep-staging datasets and one ECG dataset. WaveKernel performs better than spline on the three sleep datasets and slightly worse on ECG. Both methods have stable accuracy across the tested sampling rates. The authors also study different geometry settings, coefficient choices, and noise conditions.

Critical strengths
  1. I think the idea is interesting, and the method is clearly explained.

The paper uses optical propagation to define the basis functions of a continuous kernel. The formulation is clear, and I can follow how the kernel is constructed. The same learned coefficients can be used at different temporal coordinates. This is a useful idea for handling signals with different sampling rates.

  1. The results on the sleep datasets are consistent.

WaveKernel improves over spline on all three sleep datasets. Appendix C also reports improvements in all five matched folds for each dataset. The methods share the main architecture and prediction head, which makes the comparison useful. The grouped data splits and validation-based checkpoint selection are also reasonable. I appreciate that the authors report the weaker ECG result and the performance drop under strong noise.

Critical weaknesses
  1. My main concern is the comparison with spline.

Appendix B.3 states that the two methods use different sampling grids and alignment settings. WaveKernel also uses a basis normalization, while an equivalent normalization is not described for spline. The number of learned values is close but different: WaveKernel uses 16 values per channel pair, while spline uses 15.

Because several settings change together, it is hard to know how much of the improvement comes from the optical basis itself. The authors acknowledge these differences, but I think a comparison with matched settings is still needed. This is important because the proposed kernel basis is the main contribution.

  1. The advantage of the optical basis over other oscillatory bases is not clear.

The paper does not include a comparison with a simple Fourier or chirp basis under similar settings and parameter budgets. Such a comparison would help explain why the optical basis is useful.

The geometry and coefficient ablations show which WaveKernel settings work better. However, they do not answer whether the optical basis is better than a standard oscillatory basis. I think this leaves an important question about the contribution.

Questions for authors
  1. Could the authors compare WaveKernel and spline using the same sampling grid, kernel length, padding, alignment, and comparable normalization and parameter budgets? Does the improvement remain after matching these settings? A focused comparison on one existing sleep dataset would help address my main concern.

  2. Have the authors tested a Fourier or chirp basis in the same architecture with a similar number of learned coefficients? If these results are available, could the authors provide them?

  3. Appendix B.3 states that the CKConv implementation scales both the convolution and its bias by the inverse sampling rate. Could the authors explain this choice? Does the large accuracy drop across sampling rates remain when the bias is not scaled in this way?

3 comments
AuthorsOpenReview ↗

Matched Kernel-Basis Comparison and CKConv Bias Scaling

We thank the reviewer for highlighting the differences in experimental settings between WaveKernel and the original spline baseline. We agree that matched experimental settings are important for a fair assessment of the optical basis.

1. Matched Experimental Settings

Following your suggestion, we conducted a five-fold SleepEOG experiment, reported in Appendix D (Table 9) and summarized in Section 4.1. All methods use identical saved train/validation/test splits, backbone, prediction head, and training settings, with matched initialization of shared trainable parameters.

To isolate the effect of kernel parameterization, we use identical temporal sampling grids, kernel duration and length, SAME padding and alignment, unit stride, 1/fs1/f_s quadrature scaling, and sampled-basis RMS normalization with 1/81/\sqrt{8} scaling.

All five methods employ eight complex basis functions, 16 learnable real coefficients per channel pair, and one readout phase per layer, with the same bounded amplitude-phase parameterization and 337,129 trainable parameters. The spline control uses 16 natural cubic cardinal functions paired into eight complex basis functions, rather than the original 15-value spline baseline.

2. Comparison with Alternative Bases

The controlled experiment includes optical, spline, Fourier, Gabor, and chirp parameterizations. The results are:

Basis100 Hz accuracy (%)Seven-rate accuracy (%)Worst-rate drop (pp)
Optical59.84 ± 1.5559.48 ± 1.580.77
Gabor59.81 ± 1.5759.05 ± 1.261.06
Chirp59.45 ± 1.8458.94 ± 1.960.79
Natural cubic spline56.86 ± 1.4356.82 ± 1.380.30
Fourier56.72 ± 1.8956.32 ± 1.990.91

Accuracy is reported as mean ± population standard deviation across five folds from one seed, evaluated at 80, 100, 128, 200, 256, 512, and 1024 Hz. The worst-rate drop is computed relative to the 100 Hz accuracy within each fold, clipped at zero, and then averaged across folds.

The optical basis retains a 2.66-percentage-point improvement over spline under matched experimental settings, with a 3.16-point gain over Fourier. Optical achieves higher seven-rate accuracy than spline and Fourier in all five folds, and than Gabor and chirp in four of five folds. Its overall performance is comparable to Gabor and chirp.

The configurations of all alternative bases are detailed in Appendix D.

3. CKConv Bias Scaling

The CKConv baseline follows the publicly released NeurOCNN wrapper, which applies the 1/fs1/f_s factor after the affine convolution to approximate continuous convolution through discrete summation. This scales both the convolution output and bias, yielding (Sf+bold)/fs(S_f+b_{\mathrm{old}})/f_s, where SfS_f denotes the convolution output before bias addition.

Alternatively, the scaling can be applied only to the convolution term, giving Sf/fs+bnewS_f/f_s+b_{\mathrm{new}}, where bnew=bold/ftrainb_{\mathrm{new}}=b_{\mathrm{old}}/f_{\mathrm{train}} preserves the output at the training sampling rate.

We have not yet evaluated this alternative scaling scheme, so whether the reported cross-rate accuracy drop persists remains to be determined. We have clarified this implementation detail in Appendix B.3.

Dear author, thanks for you rebuttal. I would like to check one fact. The accuracy here is relative error or absolute error. From my view, the accuracy around 50 percents means not trained very well.

AuthorsOpenReview ↗

Clarification on Classification Accuracy

We thank the reviewer for the question. The reported values are standard classification accuracies, defined as the percentage of correctly classified samples, rather than relative or absolute errors.

Regarding the accuracy level, our experiments follow the NeurOCNN architecture (Kumar et al., ICML 2026), a published baseline for physiological time-series modeling. SleepEOG involves five-class sleep staging using a single EOG channel, which is challenging because this modality provides limited information for distinguishing sleep stages typically identified using multiple physiological signals. Under the same evaluation protocol, the spline-based NeurOCNN baseline achieves 55.78% accuracy, compared with 59.44% for WaveKernel (Table 1).

Furthermore, our newly added matched comparison uses the same backbone, training settings, and parameter budget across five kernel bases. WaveKernel achieves 59.48% accuracy, outperforming spline (56.82%) and Fourier (56.32%) by 2.66 and 3.16 percentage points, respectively (Table 9).

We hope these clarifications and additional experimental results help address your concerns and support a more favorable assessment of our work. Thank you again for your time and consideration.

Reviewer_aSwXOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposed to use ideas from optical propogation to parameterize a continuous convolution kernel that support sampling at different resolution. Authors tested out this idea in several physiological time-series benchmark comparing to baselines such as NeurOCNN combined with different ways to parameterize the continuous convolution kernels. The performance is un par with the state-of-the-art and NeurOCNN with spline kernel.

Critical strengths

If there is something special about the physical system from which we borrowed the kernel parameterization, then there might be some novelty. But it's not clear to me from the writing or from the results what kind of unique advantages should we look for from this optical kernel.

Critical weaknesses
  • Motivation. it's not clear to me what is the issue of the current continuous kernel parameterization and why adapting a learnable amplitude-phase modulation unit can solve such issue. It's not clear to me what kind of nice physical or mathematical properties we do look for inside the optical wave propagation.
  • The results are not super positive. Table 1 shows that the performance of most tasks are un par with NeurOCNN + Spline except for Sleep-EDF (EOG) task.
  • While the paper is evaluated on a very narrow set of tasks and baselines, the title and the writing of the paper might suggest this is a general neural network architecture for neural operator learning. It's not clear to me whether this approach can be applicable to other commonly tested tasks for operator learning such as solving PDE, which requires some way to generalize to more spatial dimensions.
Questions for authors

Would be great to get some more clarity on what kind of unique advantage should we expect from using optical field generated by a slab-type diffraction model. What kind of function can this express but the alternative parameterization could not. How do we ground this experimentally.

1 comment
AuthorsOpenReview ↗

Clarification of the Propagation-based Inductive Bias and Controlled Experimental Evidence

We thank the reviewer for highlighting the need to clarify the motivation, novelty, and scope of our work. We have revised the Introduction and added theoretical analysis and controlled experiments to address these concerns.

1. Motivation and Problem Formulation

WaveKernel investigates whether a compact, propagation-derived basis provides a useful inductive bias for continuous temporal kernels under a fixed parameter budget. Rather than claiming a strictly larger function space, our contribution lies in the optical propagation-based parameterization, which may enable more efficient representations of certain oscillatory kernels. Learnable amplitude-phase modulation further enables task-specific adaptation within this structured function family.

2. Inductive Bias from Optical Propagation

The key distinction is the optical propagation-based modulation of amplitude and phase. From Eq. (2), the phase curvature of an optical basis function is

ϕn′′(τ)=κγ2d2[d2+(γτ−yn)2]3/2.\phi_n''(\tau)=\frac{\kappa\gamma^2d^2}{[d^2+(\gamma\tau-y_n)^2]^{3/2}}.

Unlike Fourier and conventional Gabor atoms with constant carrier frequency, the optical response couples a spatially varying envelope with nonlinear phase modulation. Unlike fixed quadratic chirps, its phase curvature generally varies with position. Compared with spline bases, it directly incorporates oscillatory structure.

Thus, the proposed novelty is a optical propagation-based inductive bias, rather than a strictly larger function space. Alternative basis functions can approximate similar responses, but their representational efficiency may differ under a fixed coefficient budget. We clarify this distinction in the revised Introduction and Appendix A.5.

3. Controlled Comparison of Kernel Parameterizations

We added a five-fold SleepEOG experiment comparing optical, spline, Fourier, Gabor, and chirp bases under matched conditions, including identical backbone, temporal grids, training protocol, and total trainable parameter count (337,129).

Averaged over seven sampling rates, the optical basis achieves 59.48 ± 1.58% accuracy, compared with 56.82 ± 1.38% for natural cubic spline, 56.32 ± 1.99% for Fourier, 59.05 ± 1.26% for Gabor, and 58.94 ± 1.96% for chirp.

These results demonstrate gains of 2.66 and 3.16 percentage points over spline and Fourier, respectively, while showing comparable performance to Gabor and chirp. These results provide empirical support for the effectiveness of the propagation-based parameterization under matched experimental conditions.

4. Scope and Limitations

We have clarified that the current contribution focuses on physiological time-series classification. The original results show improvements over the spline baseline on three sleep-staging datasets, with slightly lower performance on ECG. Extensions to higher-dimensional tasks and PDE operator learning remain directions for future investigation.

We have revised the Introduction and Limitations accordingly, with additional theoretical analysis and controlled experimental results provided in Appendix A.5 and Section 4.1 (Appendix D), respectively.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Response and Summary of Revisions

We thank all reviewers for their constructive feedback. We have revised the manuscript to clarify WaveKernel's motivation and contribution, strengthen the theoretical analysis, and provide additional experimental evidence. The main revisions are summarized below.

1. Motivation and Contribution.

We have refined the problem formulation to highlight the inductive bias introduced by optical propagation-based kernel parameterization under a fixed parameter budget, with a focus on physiological time-series classification.

2. Theoretical Analysis.

We have expanded the mathematical analysis of optical basis functions, particularly their amplitude and phase characteristics, and clarified their distinctions from spline, Fourier, Gabor, and chirp bases (Appendix A.5).

3. Controlled Experimental Comparison.

We have added a five-fold SleepEOG experiment comparing five kernel parameterizations under matched experimental settings, with 337,129 trainable parameters for each method. WaveKernel achieves 59.48 ± 1.58% accuracy averaged over seven sampling rates, outperforming spline and Fourier by 2.66 and 3.16 percentage points, respectively, while showing comparable performance to Gabor and chirp (Section 4.1 and Appendix D, Table 9).

4. Experimental Clarifications and Scope.

We have clarified the cross-rate evaluation protocol and refined the scope of the contribution to physiological time-series classification, with corresponding updates to the experimental details and Limitations section.

These revisions are incorporated into the updated manuscript, with detailed responses provided to each reviewer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.