New developments of the Z-EDM algorithm

In this paper we address some open questions on the recently proposed Zero-Error Density Maximization algorithm for MLP training. We propose a new version of the cost function that solves a training problem encountered in previous work and prove that the use of a nonparametric density estimator preserves the optimal solution. Some experiments are reported comparing this cost function to the usual mean-square error and cross entropy cost functions.

• Type: Conference
• Year: 2006
• Where: ISDA
• Authors: Luís M. Silva, Joaquim Marques de Sá, Luís A. Alexandre
• Keywords: Cost Functions, Density Maximization Algorithm, ISDA 2006, Nonparametric Density Estimator, Operating Systems