Abstract
In this paper we propose and discuss several new approaches to noise-resistant training of multilayer perceptron neural networks. Two groups of approaches: input ones, based on instance selection and outlier detection, and output ones, based on modified robust error objective functions, are presented and compared. In addition we compare them to some known methods. The experimental evaluation of the methods on classification and regression tasks and comparison of their performances for different amounts of noise in the training data, proves the effectiveness of the proposed approaches.
Highlights
Artificial neural networks are one of the most popular models applied to predictive analysis
The instance selection algorithm we used is Generalized Edited Nearest Neighbor (GenENN) (Kordos et al 2013), which we developed from the ENN algorithm (Wilson 1972)
The multilayer perceptrons (MLP) neural network is trained on the test subsets using the Variable Step Search (VSS) algorithm with one a an appropriate error function (MSE, ILMedS, least trimmed absolute value (LTA), trapezoid error function (TEF)) and a possible error weighting if k-NN Global Anomaly Score was used in the preceding step or discarding some of the instances if the generalized ENN was used in the preceding step
Summary
Artificial neural networks are one of the most popular models applied to predictive analysis. The first group, so-called robust learning algorithms, is based mainly on novel error performance measures, derived from robust statistical estimators. In these approaches the training data is left in its original, potentially contaminated, state, but the network training process is modified. From that group we presented in Kordos and Rusiecki (2013) the LMLS (Least Mean Log Squares), MAE (Median of Absolute Errors), MIF (Median Neuron Input) and MedSum performance measures. The second group is based on instance selection and outlier detection methods In this case, the training data are reduced or corrected, to remove the impact of outliers. The training data are reduced or corrected, to remove the impact of outliers In the last section, we conclude the work and discuss in which cases each method should be used and in which conditions the methods proposed by us display their superiority
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have