A1157
Title: From robust neural networks toward robust nonlinear quantile estimation
Authors: Jan Kalina - The Czech Academy of Sciences, Institute of Information Theory and Automation (Czech Republic) [presenting]
Abstract: Regression quantiles provide a flexible framework for modeling the conditional distribution of a response variable by estimating different parts of its distribution, thereby offering valuable insights into the relationship between predictors and outcomes. However, existing nonlinear regression quantile methods may be sensitive to the presence of severe outliers in the data. Robust versions of neural networks are investigated, including a sequential outlier detection procedure based on sequential example selection for robust neural networks. Robust quantile estimators for nonlinear regression are introduced, inspired by least weighted squares regression. The proposed quantiles assign implicit weights to individual samples and are specifically tailored for multilayer perceptrons, radial basis function networks, and regularized networks. Numerical experiments demonstrate that the robust quantiles improve generalization and outlier resistance, with simulations confirming that the proposed method outperforms traditional non-robust quantiles.
