Removal of catastrophic noise in hetero-associative training samples

In many applications, sensor failures, recording errors, and source limitations can affect data collection to the extent that a significant proportion of the training set consists of "malicious" training vectors. We present a method for detecting malicious vectors in hetero-associative training samples. We propose a general metric to quantify maliciousness and investigate four methods for dealing with the problem. We present an algorithm which permits the incremental augmentation of the noise-free part of the data set, and show that it is in general superior to other possible techniques. In particular, we show that the algorithm yields faster convergence and better generalisation for small percentages of catastrophic noise in the training sample.