Structural Kolmogorov-Arnold Convolutions: Learnable Function on the Values or the Filter Shape as Parameter-Efficient Alternative to Per-Edge Convolutional KANs
arXiv:2606.24371v1 Announce Type: cross Abstract: Convolutional Kolmogorov--Arnold Networks (KANs) replace the fixed weights of a convolutional kernel with learnable univariate functions. The dominant formulation attaches one such function to every kernel entry and lets it act on pixel values, expressive but parameter-heavy and prone to overfitting. We argue that the learnable functions are better placed in the \emph{structure} of the convolution than on each edge, and we organise the design space along a single axis: whether the function acts on the pixel \emph{values} or on the filter \emph{
The continuous evolution of neural network architectures drives research into more efficient and expressive models, particularly as computational demands for current architectures increase.
This research introduces a parameter-efficient alternative to existing Convolutional KANs, potentially leading to more scalable and robust AI models, especially in resource-constrained environments.
The proposed 'Structural Kolmogorov-Arnold Convolutions' shift the complexity from individual kernel entries to the overall convolutional structure, offering a new approach to designing learnable functions in KANs.
- · AI researchers
- · Hardware manufacturers (for more efficient models)
- · Developers needing parameter-efficient models
- · Architectures with high parameter counts or overfitting issues
Improved efficiency and generalization in convolutional neural networks using KANs.
Faster development and deployment of complex AI models due to reduced computational overhead and data requirements.
Broader accessibility of advanced AI capabilities to organizations with limited compute resources, fostering wider innovation.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI