Elevating Vision Transformer Performance: The KAN-VIT Model for Image Recognition

In this paper, we propose an innovative image recognition model called KAN-VIT, which applies the Kolmogorov-Arnold Networks (KANs) architecture to the Vision Transformer (VIT), replacing the traditional multilayer perceptron (MLP) layers and the MLP head at the output. KANs are neural networks based on the Kolmogorov-Arnold representation theorem, which introduces learnable activation functions between nodes, parameterized by spline functions, thereby providing higher accuracy and interpretability. Based on this, the VIT model was improved by replacing the original MLP layers with KANs, and experiments were conducted on the CIFAR-10, Oxford Flowers, and MNIST datasets. The results show that the improved model has demonstrated superior performance on the test datasets. We also analyzed the interpretability of KANs, and the results show that the improved model not only has improved performance but also has advantages in the visualization and interpretability of the model.