Arbeitspapiere der GMD 735

The problem of selecting critical examples for training neural networks is discussed. Based on Bayesian statistics and information theory we derive a measure of criticality and present a computationally efficient method for selecting critical examples. This leads to a new type of learning scheme in which the network is trained incrementally on an increasing number of self-selected examples, instead of on all the available data at the outset. Experimental results show that the selective incremental learning method finds a critical subset of the given examples and thereby achieves considerable improvement in training speed and generalization performance if a large amount of data is known in advance.