DATCNN: A novel CNN network with all the advantages of KAN while offering greater flexibility.

Abstract

In recent years, the interpretability of artificial intelligence models has been increasingly valued. Kolmogorov-Arnold Network (KAN) is a novel neural network architecture that supports symbolic regression and offers good interpretability. Derivative works of KAN not only provide interpretability but also demonstrate good performance in common tasks such as image segmentation and time series prediction. However, KANs suffer from slow convergence and difficult training processes. To address these limitations, we propose the Dense Affine Transformation Convolutional Neural Network (DATCNN). This novel CNN-based architecture preserves the interpretability and functional representation capacity of KANs while offering enhanced flexibility and compatibility with established CNN theory and training heuristics. Experimental results across multiple tasks, including function fitting, image classification, and natural language processing, demonstrate that DATCNN achieves faster training speeds and superior performance, highlighting its potential as an efficient alternative to KANs in theoretical and practical settings.