Kolmogorov-Arnold Networks with Trainable Activation Functions for Data Regression and Classification
Kuan‐Lin Chen, Jian–Jiun Ding · 2025
Deep learning has become a widely utilized approach for a variety of applications. Some networks, such as Multilayer Perceptrons (MLPs), apply input feature data to compute regression and classification results. In recent years, many architectures, such as transformers, incorporate MLPs for backend computations and achieve notable success across a range of applications. In 2024, a novel network architecture, named the Kolmogorov-Arnold Network (KAN), has been introduced. Like MLPs, KANs share similarities in structure, but are grounded in the Kolmogorov-Arnold representation theorem. Unlike MLPs, which apply fixed activation functions at each neuron, KANs use trainable activation functions directly, making the model require fewer network layers and neurons for training. In addition, we conducted a comparative analysis using other models such as the radial basis function (RBF) network. This analysis aims to the training performance and compare the efficiency of KANs and MLPs on commonly used machine learning datasets. The results show that KANs outperform MLPs in regression and classification tasks.