Robust Loss Functions for RBF Networks and Their Derivatives

Authors: Jan Kalina, Kristýna Dvořáková
Publication year: 2027
Publication type: Konferenční příspěvek (mezinárodní konf.)
Periodical: Artificial Neural Networks and Machine Learning – ICANN 2026 Proceedings. Part I
Projects: LEDNeCo: Low Energy Deep Neurocomputing
Bibliographic citation: KALINA, Jan, DVOŘÁKOVÁ, Kristýna. Robust Loss Functions for RBF Networks and Their Derivatives. In: PASA, L., NAVARIN, N., LINTAS, A., VILLA, A. E. P., TETKO, I. V., MICHELI, A., TORTORELLA, D., POLATO, M., eds. Artificial Neural Networks and Machine Learning – ICANN 2026 Proceedings. Part I. Cham: Springer, 2027, S. 683-694. Lecture Notes in Computer Science, 17089. ISBN 978-3-032-38397-6
Dostupné z: doi:10.1007/978-3-032-38398-3_56.
DOI: 10.1007/978-3-032-38398-3_56
ID: 654353
Online record: View online record

Abstract

This paper explores highly robust methods for training Radial Basis Function (RBF) networks, focusing on derivatives of their loss functions. RBF networks are well-known for their ability to model complex nonlinear relationships, but their sensitivity to outliers and anomalies poses a challenge in real-world applications. To address this, we recall and build upon several recently proposed highly robust versions of RBF networks, which are inspired by robust parameter estimation techniques used in linear regression. We provide a detailed derivation of derivatives of the fitted values and loss functions, which are critical for optimizing the network and enhancing its robustness to noisy data. Furthermore, the paper discusses the significance of these derivatives in the context of training and model performance, emphasizing their role in improving the network’s stability and robustness. The results presented here contribute to the development of more reliable and efficient RBF networks, especially for tasks involving complex and noisy datasets.