Structural Kolmogorov-Arnold Convolutions: Learnable Function on the Values or the Filter Shape as Parameter-Efficient Alternative to Per-Edge Convolutional KANs

What Happened

A new arXiv preprint introduces "Structural Kolmogorov-Arnold Convolutions" (S-KAC), a variant of convolutional Kolmogorov-Arnold Networks that rethinks how learnable functions are applied in convolutional layers. The core innovation is a shift in focus: instead of attaching a separate learnable univariate function to every pixel value in a kernel window (as in standard per-entry KAN convolutions), S-KAC applies learnable functions to the filter shape itself—the spatial arrangement and relative positions of kernel entries—while keeping the value-based transformations simpler.

This structural approach treats the convolution kernel not as a fixed grid of weights with learnable value mappings, but as a parameterized geometric template where the positions and interactions of kernel elements become the primary learnable components. The authors demonstrate that this method achieves comparable or superior performance on standard benchmarks while using significantly fewer parameters than per-entry KAN convolutions.

Why It Matters

The KAN architecture has been a hot topic in 2025, promising more interpretable and parameter-efficient alternatives to MLPs. However, early convolutional KAN implementations suffered from a critical flaw: attaching a learnable function to every pixel value in a kernel window leads to massive parameter bloat. For a 3×3 kernel, that’s nine separate learned functions instead of nine scalar weights—a 9x increase in complexity per kernel.

S-KAC addresses this by recognizing that spatial structure in convolutions is as important as value transformations. By learning filter shape as a function, the model captures geometric invariances and spatial patterns without needing to parameterize every pixel-value interaction. This is particularly relevant for tasks where local spatial relationships (edges, textures, patterns) dominate over precise pixel intensities—which describes most computer vision tasks.

For AI practitioners, this matters because it offers a concrete path to deploying KAN-based architectures in vision models without the prohibitive computational cost that previously limited them to small-scale experiments. The parameter efficiency gains could enable deeper or wider KAN convolutions within existing memory budgets.

Implications for AI Practitioners

Architecture Design: Practitioners building vision models should consider S-KAC as a drop-in replacement for standard convolutions in resource-constrained settings. The method is particularly promising for mobile or edge deployment where parameter count is critical. Training Dynamics: The structural approach may converge differently than value-based KANs. Practitioners should expect different learning rate schedules and initialization strategies, as the optimization landscape for shape parameters differs from value parameters. Interpretability Trade-offs: While KANs promised interpretability through learned univariate functions, S-KAC’s structural focus may produce less immediately interpretable filters. The trade-off between parameter efficiency and model transparency needs careful evaluation for applications requiring explainability. Hardware Optimization: Current GPU hardware is optimized for fixed-shape convolutions. S-KAC’s dynamic filter shapes may require custom kernels or compilation passes to achieve peak throughput, potentially limiting immediate adoption in production pipelines.

Key Takeaways

• Structural KAN convolutions learn filter shape as a function rather than attaching separate functions to each pixel value, achieving parameter efficiency without sacrificing performance.
• This approach directly addresses the parameter bloat problem that previously limited convolutional KANs to small-scale experiments.
• Practitioners should evaluate S-KAC for vision tasks where spatial relationships dominate over precise pixel intensities, particularly in resource-constrained deployment scenarios.
• Adoption may require custom hardware optimization and careful tuning of training hyperparameters, as the optimization dynamics differ from standard convolutions.