A New Cost Function to Build MLPs by Means of Regularized Boosting

M. Lázaro-Gredilla, J. Madrid-Sánchez, and A.R. Figueiras-Vidal (Spain)


Boosting, Cost Functions, Parzen Models.


In this paper we propose a novel cost function to train a standard SLP, which proves to perform similarly to a linear SVM, currently considered as one of the best lin ear discriminants. Then we show how we can use regu larized boosting to construct a conventional MLP which does not suffer from overfitting, and needs no adjust ment in the number of hidden units. This MLP consis tently outperforms carefully constructed regularized and non-regularizedMLPs trained using backpropagation, non linear SVMs and sometimes even MLP ensembles (con structed with Real Adaboost-like algorithms), with the ad vantage of a much simpler structure. To define each of these MLP models, just two parameters are needed, the booster regularization parameter and λ, which accounts for regularization in the learners.

Important Links:

Go Back