Sensitivity of radial-basis networks to single-example decision classes

Some applications require classifiers that can learn classification information from very few examples per a decision class. In such applications, it is difficult or expensive to gather more than one example per a class. This paper provides an analysis of the performance of the radial-basis neural networks when trained on one example per decision class. Radial-basis neural networks are widely used in many applications. The empirical analysis investigates the recognition accuracy of these classifiers in surrounding area of known examples. This is done by changing number of attribute values, or changing a single value with different margins. The results show a strong relationship between the distance of the testing records to the training ones and the predictive accuracy. Similar distances caused by changing different attributes provide same error rate. The results indicate that distance-base classifiers (such as k-nearest neighbor) are much better classifiers than radial-basis networks with very few training data.