pRAM n-tuple Classifier - a new architecture of probabilistic RAM neurons for classification problems
Paulo J. L. Adeodato, Rosalvo Ferreira de Oliveira Neto · 2010
RAM-based neural networks have been in the market long before the MultiLayer Perceptrons but their developments have been concentrated in a few periods and research groups. Recently, one of the winners of the PAKDD 2007 Competition used a variation of one of such architectures in their data mining solution. Now, that approach for binary decision problems has been refined and is presented in this paper as the pRAM n-tuple Classifier. It consists of a single layer neural network inspired by the n-tuple classifier but with advantages in several aspects: 1) it does not need special binary encoding for continuous variables, 2) it generalizes smoothly, and 3) its performance improves with the increase in the amount of training examples without saturating. Experimental comparison carried out on the Proben1 benchmark problems against MultiLayer Perceptrons shows that there is no statistically significant difference between them. The pRAM n-tuple Classifier also has the advantage of using all modeling data for training once that it does not require a validation set, which is particularly important for problems with small data sets.