Feedforward Neural Networks for Nonparametric Regression

Abstract

Feed forward neural networks (FFNN) with anunconstrained random number of hidden neuronsdefine flexible non-parametric regression models.In Mueller and Rios Insua (1998) we have arguedthat variable architecture models with random sizehidden layer significantly reduce posteriormultimodality typical for posterior distributionsin neural network models. In this chapter we reviewthe model proposed in Mueller and Rios Insua (1998)and extend it to a non-parametric model by allowingunconstrained size of the hidden layer. This ismade possible by introducing a Markov chain MonteCarlo posterior simulation scheme using reversiblejump (Green 1995) steps to move between differentsize architectures.