A0240
Title: Bayesian Kolmogorov-Arnold neural model
Authors: Myung Won Lee - University of Edinburgh (United Kingdom) [presenting]
Miguel de Carvalho - University of Edinburgh and Universidade de Aveiro (Portugal)
Brian Reich - North Carolina State University (United States)
Abstract: The aim is to present a novel statistical framework for Kolmogorov-Arnold neural networks (KAN), redefining them as a neural extension of generalized additive models. The canonical version of the model employs a doubly-additive approach in a three-layer configuration, while the deep version extends this to a deep additive model. The approach then motivates a novel picking-and-pruning prior, specifically designed for KAN architectures. This prior facilitates group-wise regularization of the spline coefficients governing the learnable activation functions, applied simultaneously to the inner and outer layers. Thus, the method not only performs variable selection but also streamlines the dense network, leveraging Bayesian inference to enhance interpretability for both prediction and classification tasks. The proposed method is validated with simulation studies on artificial data. The applicability of the model is further investigated on the chemical exposure and inflammation data.
