PoolKANNeXt: A new pooling-based Kolmogorov Arnold convolutional neural network

Abstract

The Kolmogorov-Arnold Network (KAN) is a new-generation neural network. It provides an alternative to multilayer perceptrons (MLPs). PoolFormer showed that pooling alone can mix features efficiently. We propose PoolKANNeXt, a CNN that merges the KAN structure with pooling-based feature mixing. This design targets high accuracy with a low number of the learnable parameters. We evaluated PoolKANNeXt on six image datasets, five biomedical and one general (CIFAR-10). The model comprises four stages. In the stem stage, a ConvNeXt-style patchify block converts each 224 x 224 x 3 image into a 56 x 56 x 96 tensor. In the main stage, average pooling first mixes local features; then two parallel 3 x 3 convolutions-one followed by GELU and the other by Swish-extract complementary representations, and a 1 x 1 convolution scales the combined output and adds it back to the input via a residual connection. In the downsampling stage, strided 3 x 3 convolutions halve spatial dimensions and double the channel count. In the output stage, global average pooling produces a feature vector that feeds into a softmax classifier. PoolKANNeXt achieved over 90 % accuracy on all datasets and reached 99.43 % on CIFAR-10, ranking among the top five models on that benchmark. PoolKANNeXt offers a lightweight yet powerful architecture. Its innovative combination of pooling and dual-activation KAN blocks yields strong performance across diverse tasks. The design is scalable and adaptable to larger or more complex datasets.