Query based learning in a multilayered perceptron in the presence of data jitter

Abstract

Stochastically perturbed feature data is said to be jittered. Jittered data has a convolutional smoothing effect in the classification (or regression) space. Parametric knowledge of the jitter can be used to perturb the training cost function of the neural network so that more efficient training can be performed. The improvement is more striking when the addended cost function is used in a query based learning procedure.<>