[Comparison of polarity profiles by novelty filters]

The present study compares the ability of humans with the ability of neural networks to determine similarities between several graphic patterns. The simulated neural network was designed as a novelty filter. The stimulus material consisted of 8 different polarity profiles which were judged for their similarity under different conditions: (1) simultaneous presentation and (2) serial presentation. The judgments were collected by using the method of paired comparisons. The profile patterns were fed into the neural network to calculate novelty (similarity) values. As a further criterion for the correspondence of the profiles, a correlation coefficient (Q-coefficient) was taken. The comparison of the resulting similarity matrices shows that the performance of the novelty filter is in a good competition with the performance of a human observer. Furthermore, in both conditions the correspondence between the similarity-judgments of the subjects and the novelty values was higher than the correspondence between the judgments and the Q-coefficient. Despite this advantage the standardization of the novelty values raises some methodological problems.