acceptodds
Under review as a conference paper at ICLR 2027

Core-KAN: Continuous Vision Kernels with Kolmogorov-Arnold Networks

Abstract

Conventional convolutional kernels are typically defined on fixed discrete grids, limiting their ability to accommodate heterogeneous local structures. Existing adaptive operators improve flexibility, but often couple geometric scale variation with content-dependent filtering, while incurring high computational cost from per-location kernel generation. To decouple geometric scale adaptation from content-dependent filtering while avoiding expensive per-location kernel generation, we propose Continuous Relative-scale KAN (Core-KAN), a relative-scale-conditioned continuous convolution operator. In detail, Core-KAN maps input features into a compact latent basis space and uses a lightweight scale controller to predict local scales relative to an exponential moving average reference. A KAN-based generator represents depth-wise kernel bases as continuous coordinate functions, allowing the operator to synthesize spatial filters at arbitrary resolutions rather than being confined to a fixed lattice. Instead of synthesizing independent kernels at every location, it constructs a compact bank of scale-conditioned kernel responses and interpolates them according to the predicted local scale map. An independent mixing controller further combines the interpolated basis responses based on local content, explicitly decoupling geometric scale adaptation from content-dependent filtering. Together with lightweight pointwise projections, this design forms a low-rank dynamic convolution that scales efficiently with kernel size and can be readily integrated into hierarchical vision backbones. Extensive experiments across three representative computer vision tasks demonstrate that Core-KAN consistently outperforms strong convolutional and dynamic-kernel baselines while introducing only marginal parameter and computational overhead. Core-KAN provides an efficient and general framework for continuous, scale-adaptive convolution across diverse vision tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Continuous convolution, Scale-adaptive kernels, Dynamic convolution, Geometric-content decoupling, Efficient vision backbones

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_C56tOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Core-KAN, a relative-scale-conditioned continuous convolution operator. In detail, Core-KAN maps input features into a compact latent basis space and uses a lightweight scale controller to predict local scales relative to an operator-specific EMA reference, while a shared KAN-based generator represents depth-wise kernel bases as continuous coordinate functions on a fixed sampling grid. Instead of synthesizing independent kernels at every spatial location, it constructs a compact bank of scale-conditioned kernel responses at M supports and interpolates them according to the predicted local scale map, while an independent mixing controller combines the interpolated basis responses based on local content. Extensive evaluations on ImageNet-1K, COCO, and ADE20K demonstrate that Core-KAN consistently outperforms strong convolutional and dynamic-kernel baselines across classification, object detection, instance segmentation, and semantic segmentation tasks while introducing only marginal parameter overhead.

Critical strengths
  • Efficient Support-and-Interpolate Readout with Exact Local-Kernel Equivalence: Core-KAN constructs a compact bank of scale-conditioned kernel responses at M relative supports and interpolates them according to the predicted local scale map. By linearity in the kernel weights, this readout exactly applies the interpolated local kernel without materializing a separate kernel tensor at every position. Together with lightweight pointwise projections that confine spatially adaptive computation to an N-dimensional low-rank basis space, this design reduces generator coordinate queries.

  • Explicit Decoupling of Geometric Scale Adaptation and Content-Dependent Filtering: Core-KAN decouples spatial adaptation through a dual-control mechanism: a scale controller adapts kernel profiles through relative-scale coordinate queries calibrated by a non-learnable EMA reference, while an independent mixing controller combines basis responses according to local content. Kernel evolution and controller diagnostics demonstrate that the KAN learns smooth, basis-specific PCA trajectories rather than arbitrary discrete lookup tables, with locations closer to ground-truth boundaries exhibiting larger scale offsets and more concentrated basis mixtures.

  • Strong Parameter Efficiency and Consistent Empirical Gains Across Vision Benchmarks: Integrated into ResNet-50 on ImageNet-1K, Core-KAN achieves 81.45% top-1 accuracy with 26.61M parameters, delivering a 3.84% relative improvement over ResNet-50 with a parameter overhead of only 4.11%, and outperforming KernelWarehouse while using 73.92% fewer parameters.

Critical weaknesses
  • Complete Absence of Task-Level Ablations and Unverified Necessity of the 2D KAN Generator: The paper provides no task-level ablation studies to justify its core design choices, such as replacing the KAN generator with a parameter-matched MLP, removing the EMA reference μα\mu_\alpha, or varying the number of supports MM and bases NN. Furthermore, as acknowledged in Appendix C.4, the reported diagnostics are descriptive and do not isolate the causal contribution of either the scale controller or the mixing controller to task performance. Moreover, although the abstract claims that Core-KAN synthesizes spatial filters at arbitrary resolutions, all experiments fix the sampling grid to 3×33\times 3 with fixed sampling offsets δr\delta_r; consequently, querying Φθ(sqr)\Phi_\theta(s q_r) only evaluates 1D rays along the scalar scale ss without altering the physical receptive-field extent, leaving the necessity of a 2D continuous KAN representation unsubstantiated.

  • Omission of Computational Overhead Metrics (FLOPs, Memory, Latency) and Unaligned Baseline Comparisons: While the paper claims computational efficiency and marginal overhead, Tables 1–4 report only parameter counts (Params (M)) and omit FLOPs/MACs, peak GPU memory, and actual inference latency/throughput. And all experiments are restricted to a single backbone (ResNet-50), and Table 1 compares Core-KAN trained with an aggressive 300-epoch recipe against baselines with differing training recipes, which compromises the fairness of the ImageNet-1K comparison.

Questions for authors
  1. Component & Generator Ablations: Can you provide task-level ablation results that isolate: (a) the KAN generator versus a parameter-matched MLP generator, (b) the individual causal contributions of the scale controller, the EMA reference, and the mixing controller, and (c) task performance under different support counts and basis dimensions?

  2. Computational Cost & Latency: What are the exact FLOPs/MACs, peak GPU memory usage, and actual inference throughput (images/sec or latency in ms on your GPU) of Core-KAN compared with ResNet-50, ODConv, KernelWarehouse, and FDConv across Tables 1~4, including the specific breakdown of the controllers and the MM-support response bank?

  3. Recipe Alignment & Arbitrary-Resolution Generalization: What is the ImageNet-1K validation accuracy of the vanilla ResNet-50 baseline when trained under the exact same recipe specified in Appendix B.3? In addition, since Core-KAN is formulated as a resolution-agnostic continuous kernel operator, how does it perform when synthesizing larger kernels or when integrated into lightweight and modern backbones?

Reviewer_dLy4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Core-KAN, a spatially adaptive convolution in which a KAN generates continuous depthwise kernel bases from scale-transformed coordinates. Separate controllers predict a local relative scale and content-dependent basis weights, while a support-and-interpolate scheme avoids generating a kernel at every location. Results on ImageNet-1K, COCO, and ADE20K show consistent gains over the reported convolutional and dynamic-kernel baselines with modest parameter overhead.

Critical strengths
  1. The operator is clearly formulated. In particular, the paper carefully distinguishes exact response-space interpolation from approximation to direct continuous-kernel evaluation and provides a reasonable conditional error analysis.

  2. The reported results are consistently strong across classification, detection, instance segmentation, and semantic segmentation, suggesting that the overall architecture is practically promising.

Critical weaknesses
  1. The central contributions are not isolated experimentally. There is no comparison of the KAN generator with an MLP or directly learned support kernels, nor ablations of the scale controller, mixing controller, EMA normalization, or their proposed decoupling. Because the inference-time operator is effectively a structured mixture of a finite fixed kernel bank, these comparisons are necessary to establish the claimed novelty and explain the gains.

  2. The efficiency and comparative evidence are incomplete. The paper reports parameters but not FLOPs, latency, throughput, or memory, despite computing multiple support responses and two controller branches. ImageNet recipes differ across methods, results are single-seed point estimates, and the exact blocks replaced by Core-KAN are not specified.

Questions for authors
  1. Under matched parameters and training, how does Core-KAN compare with an MLP coordinate generator and with directly learned support kernels?

  2. What are the accuracy contributions of the scale branch, mixing branch, EMA-relative normalization, and using separate rather than shared controllers?

  3. What are the measured FLOPs, latency, throughput, and peak memory relative to standard convolution and the strongest dynamic-convolution baselines?

Reviewer_h5jrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Core-KAN introduces spatially adaptive convolution through KAN-parameterized continuous kernel fields. It interpolates a shared bank of scale-conditioned depthwise responses in a low-rank latent space, with separate controllers for local scale and content mixing. Experiments evaluate the method on ImageNet-1K, COCO, and ADE20K, with additional interpolation analysis and controller diagnostics.

Critical strengths
  1. The operator design is coherent, and the analysis is sound under its stated assumptions. Sharing and interpolating kernel responses provides location-dependent filtering without materializing dense per-location kernels. Appendices A.2-A.4 clearly distinguish the exact interpolation identity from approximation to direct continuous queries.
  2. Core-KAN has promising downstream improvements. Under the COCO 3x schedule, Core-KAN improves over FDConv from 45.7 to 46.2 box AP and from 41.7 to 42.2 mask AP. ADE20K mIoU also improves from 43.50 to 44.19. These results suggest useful transfer across detection and segmentation.
Critical weaknesses
  1. The contributions of the core design choices are not sufficiently isolated. Table 1 explicitly mixes training recipes, and downstream models inherit their respective pretrained backbones. No task-level ablations isolate the KAN generator, EMA-relative scaling, or the separate controllers. Figures 3 and 4 establish smooth kernel evolution and interpolation fidelity, but not the accuracy benefits of these choices. Since scale changes kernel profiles on a fixed grid, smooth PCA trajectories and boundary correlations alone do not establish the claimed geometric interpretation.
  2. The efficiency and implementation of Core-KAN require further clarification. No FLOPs, latency, or peak-memory measurements are reported to substantiate the marginal-overhead claim, despite the use of two convolutional controllers and an M-fold response bank. Appendix B.1 omits the exact replacement locations. Further, the stated low-rank replacements should substantially reduce later-stage convolution parameters, making the reported increase from 25.56M to 26.61M difficult to reconcile without an architectural breakdown.
Questions for authors
  1. Under an identical training recipe, how does the KAN generator compare with a comparably sized MLP or a directly learned kernel bank using the same controller and readout architectures?
  2. What are the end-to-end latency and peak-memory costs at the default M = 8, including the controllers and response bank, compared with the baseline?
  3. Which conv2 blocks are replaced, and how is the 26.61M parameter count obtained? A standard 512-channel 3×3 convolution has approximately 2.36M weights. Using the stated N = 16 and controller width of 32, I estimate approximately 0.34M parameters for the described replacement. Could the authors reconcile this estimate with the reported total?
  4. Can a controlled object-rescaling experiment compare adaptive and fixed-scale variants and examine their scale fields? This would clarify whether local scaling provides useful geometric adaptation beyond generic content-dependent filtering.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.