Convolutional-KANs

Why we included this project

Convolutional-KANs takes Kolmogorov-Arnold Networks, the architecture that puts learnable activation functions on connections rather than fixed activations at nodes, and reworks the idea into a convolutional layer. Where a normal convolution computes a dot product between kernel and image patch, this one applies a learnable nonlinear activation per pixel, so the layer acts as a trainable nonlinear map instead of a linear filter followed by a fixed activation. The code builds on the efficient-KAN codebase, which makes it straightforward to drop one of these layers into a PyTorch pipeline and compare it against a standard Conv2d. The companion paper reports accuracy, parameter counts, and per-epoch training times, giving an honest sense of the trade-offs: better expressiveness per parameter, but slower training and more parameters per kernel. Treat it as a research sandbox for testing alternatives to classic convolutions, not as production-ready software.