Active Learning Using a Constructive Neural Network Algorithm

摘要: Constructive neural network algorithms suffer severely from overfitting noisy datasets as, in general, they learn the set of examples until zero error is achieved. We introduce this work a method for detect and filter using recently proposed constructive algorithm. The works by exploiting fact that are harder to be learnt, needing larger number synaptic weight modifications than normal examples. Different tests carried out, both with controlled experiments real benchmark datasets, showing effectiveness approach.